Deep Learning Models and Social Engineering Dynamics in Insider Threat Detection: A Systematic Review

Deep Learning Models and Social Engineering Dynamics in Insider Threat Detection: A Systematic Review | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Systematic Review Deep Learning Models and Social Engineering Dynamics in Insider Threat Detection: A Systematic Review Ishara Barhoson Galadima, Norafida Bte Ithnin, Nur Haliza Abdulwahab, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7396895/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The exponential expansion of the global digital ecosystem has significantly increased organizational vulnerability to sophisticated insider threat attack vectors. Although Machine Learning and Deep Learning models have improved anomaly detection techniques, a critical gap remains in addressing insider threats influenced by internal social engineering. In particular, Reverse Social Engineering, where malicious insiders manipulate unintentional or innocent colleagues, poses an emerging and underexplored threat. This study systematically reviews forty-nine peer-reviewed articles published between 2015 and April 2025, using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses methodology to investigate current deep learning approaches for insider threat detection. The review highlights a reliance on sequential models such as Long Short-Term Memory and Gated Recurrent Unit algorithms, attention-based transformer models, and graph neural networks. These techniques demonstrate effectiveness in identifying behavioral anomalies and system misuse but fail to detect trust manipulation and social exploitation. Additionally, commonly used datasets, including the Computer Emergency Response Team Insider Threat Dataset from Carnegie Mellon University, DARPA1999, and Enron, do not adequately represent realistic social engineering scenarios, thereby limiting the ability of detection models to address human-driven threats. Traditional evaluation metrics, including Precision, Recall, and F1 Score, also fall short in assessing the contextual and behavioral dimensions of insider threats. This review emphasizes the urgent need for adaptive, context aware and behavior-aware detection frameworks, enriched datasets that incorporate social dynamics, and evaluation models that account for cognitive influence. Addressing these overlooked dimensions is essential for advancing organizational cybersecurity resilience against evolving insider threat landscapes. Insider threat detection social engineering Deep learning Behavioral cybersecurity social manipulation Human factor Systematic review Computer Emergency Response Team Insider Threat Dataset Explainable artificial intelligence Cybersecurity strategy Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1. Introduction The contemporary digital economy and the rapid evolution of cybersecurity attacks on critical sectors in the global economy ecosystem have profoundly influenced the development of Artificial Intelligence (AI)-driven detection models (Johnson and Verdicchio, 2024). As enterprises embrace digitization, insider threats have become an increasingly critical cybersecurity concern, particularly as attackers shift from purely technical exploits to exploiting human vulnerabilities (Han et al., 2023; Still and Cain, 2020). Human factors in cybersecurity are central to shaping security assurance across all domains. Therefore, the insider threat landscape is now characterized by a challenging dynamic, multifaceted risks factors, where malicious insiders and compromised employees through normal organisational standard business process or procedure can interact, influence and induced behavioural changes with sophisticated technical infrastructures to cause organizational harm (Alzaabi and Mehmood, 2024a). This phenomenon underscores the urgent need to rethink how organizations detect, mitigate, and respond to insider threats in a digital system. The rise of insider threats is further complicated by this increasing sociotechnical integration, where malicious insiders can manipulate innocent employees into performing harmful actions based on organisational behaviour unknowingly (Handri et al., 2024). Although traditional insider threat models primarily focus on monitoring privileged access or detecting anomalies, they often overlook the more subtle, socially driven manipulations that weaponize human trust (Cramer et al., 2025; Tundis et al., 2021). As the sophistication of social engineering attacks grows, bypassing technical safeguards by exploiting psychological vulnerabilities, the need for human-factor-centric detection strategies becomes increasingly paramount (Matsuda et al., 2020). However, technological advancement presents a paradox. On one hand, AI models, behavioural analytics, and deep learning techniques equip organizations with powerful detection capabilities. On the other hand, these same advances are leveraged by malicious insiders to orchestrate complex, stealthy, and socially engineered attacks (Kim et al., 2022). This duality has resulted in insider threats that are more persistent, context-aware, and difficult to detect using conventional static models. Emerging technologies such as Generative AI, personalized phishing campaigns, and deepfake communications further exacerbate these risks, illustrating the urgent necessity for dynamic, adaptive defence mechanisms that understand both technical and human behavioural indicators. Despite growing academic and industry interest, existing research on insider threat detection demonstrates critical limitations (Golda et al., 2024). While many studies have advanced deep learning-based detection approaches, they predominantly address technical anomalies and neglect the potential threat of internally driven social engineering attacks (Javaheri et al., 2024; Ivanov et al., 2021). Most frameworks (Cramer et al., 2025; He et al., 2022; Syafitri et al., 2022a; Tundis et al., 2021) view social engineering solely as an external threat vector rather than acknowledging its insidious potential when orchestrated within organizational boundaries. Moreover, research efforts often focus on classifying malicious insider behaviour without adequately incorporating sociotechnical factors such as interpersonal manipulation, emotional exploitation, or behavioural deception (S. Yuan & Wu, 2021a). As a result, insider threats induced by internal social engineering remain critically underexplored in both conceptualization and practical mitigation (Khadka et al., 2023; Remmide et al., 2024; Gallo et al., 2024). Considering the number of various studies on cybersecurity insider threats, social engineering attacks within organizational boundaries as independent studies, suffices the necessity for a systematic comprehensive analysis of “social engineering as an insider threat attack vector” to deal with these potent evolving risks (Kamatchi and Uma, 2025a; Hijji & Alam, 2021; Scott and Kyobe, 2021; Nifakos et al., 2021; Syafitri et al., 2022b; Alharthi et al., 2020). To pave this way, our study investigates the emerging class of insider threats induced by internal social engineering and critically examines the effectiveness of deep learning-based models in detecting such threats. While prior studies have predominantly focused on technical anomalies and behavioural deviations, they have paid limited or no attention to the broader sociotechnical dimensions and organizational behavioural factors that can enable malicious insiders to influence non-malicious employees. Specifically, organizational procedural weaknesses such as trust hierarchies, role ambiguities, and policy circumventions can be systematically exploited to facilitate internal manipulation, thus amplifying insider risk. By addressing these overlooked dynamics, this study aims to bridge a critical gap in insider threat detection research, advocating for context-aware, human-centric detection models that capture the subtle socio-behavioural mechanisms underpinning modern insider threats. To the best of our knowledge, this is the most comprehensive systematic literature review that explicitly addresses “social engineering as insider threats attack vector”, considers threats arising from both technological and human factors, and offers a comparative synthesis of deep learning-based detection approaches targeting these vulnerabilities. Aiming to bridge the gaps identified in previous surveys, this review is guided by the following research questions ( RQs ): RQ1: What deep learning models have been applied to insider threat detection, and how have they evolved over time? RQ2: What datasets are commonly utilized for developing, training, and evaluating these models? RQ3: What performance evaluation metrics are used to assess the effectiveness and robustness of insider threat detection models? RQ4: What are the key challenges, limitations, and emerging future directions in applying deep learning to insider threat detection? The main contributions of the paper are summarized as follow: (i) Conducting a comprehensive investigation of 49 peer-reviewed articles to identify prevailing deep learning models, benchmark datasets, and evaluation methodologies in insider threat detection. (ii) Highlighting the significant research gap related to internal social engineering, particularly reverse social engineering, as an overlooked insider threat attack vector. (iii) Proposing a novel taxonomy of insider threat attack vectors that incorporates both malicious intent and human-factor exploitation. (iv) Synthesizing the challenges and limitations identified in current deep learning approaches, including issues of class imbalance, model interpretability, and temporal behaviour drift. (v) Recommending actionable research directions for developing adaptive, human-centric behaviourally enriched, and sociotechnical-aware deep learning models capable of detecting both malicious insiders and manipulated unintentional insiders (innocent employees) aiming to redefine insider threat paradigms and provide a roadmap for advancing detection model capable of mitigating the evolving insider threats landscape. The remainder of this paper is structured as follows: Section 2 presents a review of related studies. Section 3 outlines the research methodology. Section 4 provides the qualitative synthesis and classification of detection approaches. Section 5 introduces the conceptual architecture for context-aware insider threat detection. Section 6 summarizes the key findings, and Section 7 offers the discussion and conclusion. 2. Related Studies Studies in the independent domains of cybersecurity insider threat, social engineering and deep learning for anomaly detection are extensive; however, only a few scholarly works have systematically reviewed deep learning-based frameworks specifically for insider threat detection and even limited number have explored social engineering as an internal threat vector. (Yuan and Wu, 2021a)conducted a comprehensive survey on insider threat detection using deep learning techniques by evaluating the capabilities of deep learning architectures on insider threat detection. Their study primarily focused on how temporal, spatial, and structural behaviour patterns can be captured using deep models. They categorized insider threats into traitors, masqueraders, and unintentional insiders, and highlighted challenges related to dataset realism and model interpretability. However, their taxonomy did not include socially engineered insiders particularly those manipulated through internal deception, such as reverse social engineering. A more behaviour-centric review was presented by (Kamatchi and Uma, 2025a), who investigated multi-modal and user behaviour modeling techniques. They discussed the benefits of incorporating few-shot learning and psychological profiling into deep learning architectures, showing potential for improved detection of nuanced behavioural anomalies. Yet, their study did not examine human-factor exploitation or social manipulation, leaving a gap in understanding behaviour motivated or distorted by trust exploitation (Gayathri et al., 2024a). A systematic review by and Alzaabi & Mehmood, 2024) offered a broad evaluation of malicious insider threat detection through machine learning methods, including hybrid models like DeepTaskAPT that leverage LSTM and CNN fusion. Their analysis emphasized feature fusion and sequence profiling but remained focused on direct malicious intent, rather than scenarios where users act under manipulation or coercion (Gayathri et al., 2024a; S et al., 2023a). Addressing the challenge of class imbalance, (Al-Shehari et al., 2024) evaluated the performance of CNN-based models with oversampling techniques such as SMOTE and ADASYN on insider threat datasets. They demonstrated improved accuracy in classifying malicious behaviours in synthetic data environments, particularly using the CERT dataset. Nevertheless, their model performance was evaluated purely on activity patterns, with no consideration for contextual or psychological influences behind user behaviour. A Comprehensive Taxonomy of Social Engineering Attacks and Defense Mechanisms was addressed by (Zaoui et al., 2024), who develop comprehensive taxonomies of social engineering attacks and defense mechanisms, with a particular emphasis on identifying effective mitigation strategies to enhance cybersecurity. Although their findings are valuable in mitigating against technical threats, their survey primarily focused on external adversarial vectors, and did not account for internal social manipulation. Human factors were directly examined by (Hughes-Lartey et al., 2021), who confirmed that behavioural vulnerabilities are the most critical point of failure in information security, particularly in IoT environments (Kamatchi & Uma, 2025c). Their study validated the statistical relationship between human error and data breaches but stopped short of integrating these findings into AI-driven detection mechanisms missing the opportunity to link behavioural manipulation to deep learning-based insider threat models. (Alam and Barron, 2022;Koutsouvelis et al., 2020) offered a structured overview of DL methods for insider threat detection, classifying models by input type and detection logic. Although they acknowledged the sophistication of insider behaviours, they did not explore the role of social dynamics, coercion, or trust-based manipulation within organizational hierarchies. Additionally, semantic models such as Log2Vec (F. Liu et al., 2019) and transformer-based log analysis tools (Alzu’bi et al., 2025a) showed that pre-trained architectures could successfully model event patterns using embeddings. However, these approaches focus on event semantics, not behavioural motivation, and therefore fail to capture psychologically driven insider activity like that seen in reverse social engineering (Yuan & Wu, 2021). As indicated in Table 1 of our study, none of the existing reviews or detection frameworks comprehensively consider insider threats originating from internal social engineering tactics. While most studies treat social engineering as an external attack, our work positions it as an internal vector, where well-meaning employees are manipulated by adversaries from within the organization. Furthermore, most reviews stop at the technical or behavioural level and do not integrate sociotechnical indicators into their detection method. To the best of our knowledge, this systematic literature review (SLR) is the first to provide a systematic review of deep learning approaches for insider threat detection with a distinctive focus on threats induced through internal social engineering manipulation. Unlike prior reviews that largely concentrate on malicious insiders or technical anomalies, this study emphasizes the emerging class of induced insider threats to well-intentioned individuals who could be manipulated into harmful actions through techniques such as phishing, pretexting, and reverse social engineering. This review makes four major contributions: (1) a taxonomy of deep learning model architectures applied to insider threat detection, (2) a taxonomy of datasets with attention to behavioural and contextual representation, (3) a taxonomy of evaluation metrics and their limitations, and (4) a forward-looking taxonomy of research directions emphasizing sociotechnical and behavioural model. Additionally, the study proposes a conceptual architecture that integrates behavioural, procedural, and psychological signals with traditional log-based analytics, enabling context-aware and explainable deep learning frameworks. Overall, this work addresses a pressing gap in the field by advocating for adaptive, human-centric detection systems capable of identifying both malicious and socially engineered insider threats in modern organizational environments. Table 1: Comparison between related studies, surveys and this SLR. Fig 1.2 representing the paper selection process for our systematic literature review. It shows the proportion of studies retained at each stage from initial identification through final inclusion highlighting the screening rigor applied to ensure only the most relevant 49 studies were included. Figure 2 presents the distribution of deep learning models employed in insider threat detection studies reviewed in this research. It reveals that Long Short-Term Memory (LSTM) networks are the most commonly used individual model, appearing in approximately 13 studies, which underscores their effectiveness in capturing sequential behavioral patterns in user activity data. Convolutional Neural Networks (CNNs) follow, used in around 7 studies, reflecting their adaptability in processing structured security logs. Emerging models such as Graph Neural Networks (GNNs) and Deep Neural Networks (DNNs) appear in fewer studies, with 4 and 3 mentions respectively, suggesting growing interest in relationship modeling and layered representations. Other models, including Feedforward Neural Networks (FNN), Generative Adversarial Networks (GAN), Autoencoders, Deep Belief Networks (DBN), and Gated Recurrent Units (GRU), show more limited application. Notably, the most frequent category is the combination of LSTM with other models—appearing in 21 studies—indicating a strong trend toward hybrid and ensemble architectures aimed at enhancing detection accuracy and context-awareness. This pattern highlights the dominance of sequential modeling approaches while also revealing an underutilization of generative and unsupervised learning methods, especially for detecting subtle, socially engineered insider threats. 3. Research Methodology Our Systematic Literature Review (SLR) was conducted using the established guidelines of PRISMA 2020 protocol (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) to ensure rigor, transparency, and reproducibility. The objective was to synthesize and critically analyze deep learning-based approaches for insider threat detection published between January 2015 and April 2025. Figure 3 illustrates the distribution of datasets used in the studies reviewed for insider threat detection research. A dominant 82% of the studies relied on the CERT Insider Threat Dataset developed by Carnegie Mellon University’s Software Engineering Institute. This overwhelming reliance reflects its wide acceptance as a benchmark for insider threat detection, largely due to its structured logs, labeled activities, and availability. However, such heavy dependence also underscores a major limitation in the field, as the dataset lacks realistic social engineering scenarios and context-aware behavioral annotations. Other datasets appear in much smaller proportions. Enterprise Event Logs account for 4% of the studies, followed by Critical Infrastructure Activity Dataset, Simulated Cloud Logs, and Coded Dataset, each contributing 2%. Additional niche datasets such as Sapi Mouse Dataset, Generated Dataset, Augmented Dataset, and the Enron Email Dataset also each represent 2% of the total. This fragmented use of alternative datasets suggests a scarcity of diverse, publicly available insider threat datasets that capture nuanced behavioral patterns or socio-cognitive elements. The chart highlights a significant research gap: the field lacks representative, socially annotated datasets capable of supporting advanced detection models, especially for insider threats induced by manipulation, trust abuse, and reverse social engineering. Addressing this limitation is critical for building and validating adaptive, context-aware detection frameworks. 3.1 Review Protocol The protocol included inclusion/exclusion criteria, selection strategy, quality assessment, and a systematic approach to data extraction and synthesis using literature review matrix. 3.2 Data Sources and Search Strategy A comprehensive literature search was conducted across multiple academic databases: IEEE Xplore ScienceDirect (Elsevier) Wiley Online Library Scopus Web of Science Springer ACM To ensure comprehensive coverage of relevant studies, multiple databases were consulted, including IEEE Xplore, ScienceDirect (Elsevier), Wiley Online Library, Scopus, Web of Science, Springer, and the ACM Digital Library. The distribution of articles retrieved from these sources is presented in Fig. 4 , which shows that the majority of studies were obtained from ScienceDirect (21%) and IEEE Xplore (17%), followed closely by Wiley Online Library (16%) and Web of Science (14%). Search String ("insider threat" OR "malicious insider") AND (detection OR "user behaviour") AND ("deep learning" OR LSTM OR RNN OR "attention mechanism") 3.3 Inclusion and Exclusion Criteria To ensure rigor and relevance, clear inclusion and exclusion criteria were applied during the study selection process. Only peer-reviewed, English-language articles published between 2015 and April 2025 were considered, focusing specifically on deep learning models for insider threat detection across enterprise, network, cloud, or IoT environments. Studies were excluded if they addressed unrelated cybersecurity threats such as phishing or spam, were not accessible in full text, or lacked technical implementation and evaluation. The detailed inclusion and exclusion criteria are summarized in Table 2 . Table 2 inclusion and Exclusion Criteria Inclusion Exclusion Publication Years Studies published between 2015 and April 2025 Studies published before 2015 Language English only Non-English Focus Area Studies involving deep learning for insider threat detection Studies on unrelated cybersecurity threats (e.g., phishing, spam) Accessibility Peer-reviewed, publicly available full-text articles Abstract-only or inaccessible documents Evaluation Must include model, dataset, and performance evaluation metrics Conceptual or opinion pieces without technical implementation Application Context Studies focused on enterprise, network, cloud, or IoT insider threat Studies focusing solely on external threat detection 3.4 Study Selection Process The study selection followed a structured 4-stage PRISMA process: 1. Identification : 498 articles were retrieved using the defined search terms from selected databases. 2. Screening : 146 articles were excluded based on title relevance. 3. Eligibility : 69 articles were excluded after abstract screening; 66 were retained for full-text analysis. 4. Inclusion : 20 articles were excluded after full-text review (due to insufficient methodological detail or relevance), resulting in 49 articles for qualitative synthesis. A PRISMA flow diagram (Fig. 1 ) summarizes the selection process. 3.5 Quality Assessment To ensure the scientific rigor of included studies, each was evaluated using a 5-point checklist adapted from (Kitchenham, 2007 ): Clear statement of research objective Justification of the DL technique/model used Clarity of dataset description and source Reported evaluation methodology and reproducibility 3.6 Data Extraction and Synthesis A data extraction form was developed to capture the following information: Publication details (year, authors) Deep learning model(s) used Dataset(s) used Evaluation metrics Reported challenges, limitations, and proposed solutions Descriptive and thematic analysis techniques were employed to identify patterns, trends, and research gaps. The results are grouped according to the research questions and discussed in Section 4 . 4. Qualitative Synthesis and Detection Classification After an extensive review of the 49 peer-reviewed studies included in our qualitative synthesis (Step 4), we manually classified insider threat research into three interdependent categories: sociotechnical threats, behavioural threats, and human factor threats. This expanded taxonomy builds upon previous classifications that primarily conceptualized insider threats through technical or behavioural lenses, often overlooking the crucial roles of organizational behaviour and relational dynamics within modern digital enterprises (Kamatchi & Uma, 2025c ). Several studies exemplify the evolving approaches to insider threat detection. (He et al., 2024) proposed a double-layer detection framework combining Long Short-Term Memory (LSTM) and Extreme Gradient Boosting tree (XGBoost) for phishing detection, alongside a Bidirectional LSTM with an Attention mechanism for insider threat detection. Similarly, (Xiao et al., 2024 ) introduced a comprehensive framework integrating statistical and sequential analysis through convolutional attention and a Transformer encoder (CATE) to address prior limitations in capturing temporal relationships. However, these sequential-based approaches still face challenges, including their potential inability to capture the full complexity of user behaviour and their reliance on domain expertise for rule-based statistical models, limiting their adaptability to emerging insider threats. (Song et al., 2024b ) further addressed poor detection performance stemming from ineffective use of behavioural time information by applying deep learning-based autoencoders on behavioural logs. While their approach achieved high prediction accuracy and precision, limitations such as data imbalance, insufficient label availability, and elevated false alarm rates in hybrid methods persist. Building on this gap, and to the best of our knowledge, our systematic review comprehensively integrates sociotechnical, behavioural, and human-factor dimensions into a novel and refined taxonomy for insider threat detection placing particular emphasis on internal social engineering as an emerging, critically underexplored attack vector. 4.1. Sociotechnical Threats Our qualitative papers for sociotechnical threats where technical system vulnerabilities are exploited through human interaction represents a foundational layer in contemporary insider threat research. Many studies have concentrated on uncovering anomalies within system logs, access behaviours, privilege escalations, and user-device interactions by employing advanced deep learning models (Le & Nur Zincir-Heywood, 2019 ). Early models predominantly relied on sequential analysis, with Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs) proving particularly effective in modeling system usage over time (Yeol Kim & Kwon, 2022 ; Li et al., 2024; and Manoharan et al., 2024 ). More recently, researchers have integrated Graph Neural Networks (GNNs) to capture complex system relationships (Tian et al., 2025 ). For instance, the Digital Twin Insider Threat Detection framework (Wang and El Saddik, 2023a ) employed Digital Twins combined with Self-Attention Transformers to model digital footprints of users, identifying anomalous interactions across virtual system replicas. Similarly, the Learning Adaptive Neighbors framework by (Cai et al., 2024 ) and introduction of adaptive neighbour model to detect anomalies within relational system graphs (Hong et al., 2022 ). These approaches demonstrate an increasing sophistication in model not only individual behaviours but also inter-system communication patterns indicative of insider exploitation (Le &Zincir-Heywood, 2021 ). Despite technical advances, significant gaps remain. The majority of these models treat insider threats as purely technical anomalies, neglecting the sociotechnical reality that many technical breaches are facilitated by human-driven social engineering manipulation (Ahmadi-Assalemi et al., 2022 ). For example, a malicious insider may not overtly escalate privileges but may coerce or deceive a colleague into performing privileged actions on their behalf, actions that mimic legitimate behaviour at a system log level. Current GNN and LSTM models, while capable of detecting privilege misuse or anomalous login patterns, lack the ability to discern the social context behind these actions. Moreover, models such as Ripple2Detect (H. Liu et al., 2025 ) attempt to trace multi-step evidence of insider attacks using semantic similarity learning, yet still primarily focus on technical event chaining rather than the underlying human manipulation processes that initiate those events (Sui et al., 2022 ). Although deep learning models for sociotechnical threat detection have evolved in sophistication, from sequential anomaly detection to graph-based relational anomaly learning; they largely remain blind to insider threats initiated through social manipulation. Current architectures fail to incorporate indicators of interrelatedness of social and technical aspects of organisation procedural relational exploitation, or hierarchical trust dynamics, all of which precede many technical anomalies. Bridging this gap requires new models capable of integrating sociotechnical relational behavioural signals with technical activity streams. 4.2. Behavioural Threats Behavioural threat model focuses on identifying deviations from expected user activities, aiming to detect early signs of potential insider risk. In the reviewed literature, behavioural model emerged as a dominant approach for insider threat detection, reflecting a recognition that malicious insiders often betray themselves through subtle, progressive shifts in behaviour rather than sudden technical anomalies. Deep learning models, particularly sequence-based architectures such as Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs), were frequently employed to capture temporal dependencies within user activity logs (Ranadive et al., 2023a). Several frameworks introduced enhancements to basic sequential model to better capture complex behaviour drifts. For instance, (Song et al., 2024a ) proposed a behaviour rhythm detection model incorporating time-awareness and user adaptation, dynamically adjusting anomaly baselines based on evolving user behaviours. Similarly, (Suryotrisongko et al., 2022 ) incorporated multi-step evidence aggregation, connecting temporally dispersed anomalous actions into coherent behavioural chains. Hybrid models combining convolutional layers with sequential architectures also gained popularity. (Atosha et al., 2024 ) employed a CNN-GRU hybrid to simultaneously capture spatial (file movement patterns) and temporal (login session anomalies) features, thereby improving sensitivity to nuanced insider behaviours more holistically (R.G. et al., 2024; Tian et al., 2023 ). Despite these advances, critical limitations persist. Firstly, most behavioural models focus heavily on overt anomalies such as unusual login times, off-hours system access, or spikes in data movement. They generally neglect the cognitive and emotional dimensions of behaviour that might precede technical anomalies. For example, an insider gradually coerced or manipulated through reverse social engineering might initially exhibit changes in communication style, responsiveness, or collegial interactions none of which are captured in traditional activity logs. Secondly, while models introduced by (Burrell and Nobles, 2022 ) focus on relational aspects by adapting neighbour behaviours dynamically, they still prioritize structural system behaviours rather than human relational dynamics. Most studies also struggle with distinguishing between harmless behaviour drift such as a promotion leading to access to new systems and potential malicious drift induced by internal manipulation. Although behavioural threat detection models increasingly capture sequential and temporal irregularities, they remain technically anchored and blind to behaviour deviations arising from social manipulation, cognitive bias exploitation, or emotional leverage. Addressing this gap requires next-generation models capable of integrating behavioural sentiment analysis, cognitive behavioural profiling, and relational context awareness into temporal anomaly detection models. 4.3. Insider Human Factors and Social Engineering Challenges Despite significant advances in detection technologies, modeling human factors in insider threats especially within social engineering contexts remains a critical challenge. Human vulnerabilities, including cognitive biases, emotional responses, and social dynamics, are often exploited by malicious insiders or inadvertently triggered by unsuspecting colleagues. Yet, these dimensions remain underrepresented in most existing threat detection frameworks. Social engineering attacks, such as phishing and pretexting, are designed to manipulate psychological and social susceptibilities(Taherdoost, 2024 ). However, current insider threat models predominantly focus on technical indicators, neglecting behavioural and relational cues essential to understanding these manipulative tactics(Sedes and Degrace, 2024 ). As a result, detection models struggle to capture the full complexity of insider risks. Moreover, the unpredictable nature of human behaviour limits the efficacy of static detection techniques, calling for more proactive approaches that anticipate subtle signals of compromise(C. Liu et al., 2019 ). Integrating sentiment analysis, communication patterns, and social relational behaviour into deep learning frameworks could significantly enhance the early detection of insider threats. Recognizing that human factors are central to cybersecurity resilience, this study advocates for their systematic inclusion in insider threat detection strategies(Burrell & Nobles, 2022 ). Human factors particularly emotional vulnerability, cognitive biases, and relational trust exploitation remain critically underexplored in deep learning models for insider threat detection. Without modeling how insiders are manipulated internally through social engineering, current detection systems risk overlooking the most subtle yet devastating form of insider attacks: those perpetrated unknowingly by compromised but innocent employees. Bridging this gap demands the development of context-aware, socio-cognitively enriched models capable of detecting not only what an insider does but why they behave differently. 4.3.1 Social Engineering Social engineering is a multifaceted phenomenon encompassing physical, technical, social, and socio-technical dimensions, rendering it resistant to straightforward detection or mitigation through traditional technical protection measures. In the context of insider threats, this challenge is further amplified, as human operators, technicians, and end-users often the weakest links within organizational information systems become primary targets of manipulation rather than machines, code, or technical systems (Yasin et al., 2025 ). The evolution of digital infrastructures, from monolithic systems to interconnected graphs of loosely coupled microservices, along with the proliferation of social networks and cloud-based applications, has expanded the human attack surface (Amiri-Zarandi et al., 2023a ). Recent advances in cyber-defense technologies have shifted the attack focus from system vulnerabilities to human vulnerabilities, exploiting trust, emotions, and social relationships within organizations (Fuertes et al., 2022 ). Particularly, the widespread adoption of Bring Your Own Device (BYOD) policies and the increased reliance on third-party communication platforms in enterprise environments have broadened the range of access points available for sophisticated social engineering attacks. When these human vulnerabilities are combined with the exploitation of zero-day technical weaknesses, social engineering becomes a formidable attack vector, frequently leveraged by advanced persistent threat actors to compromise even highly secured financial and enterprise systems(Kamruzzaman et al., 2023 ). Given these developments, internal social engineering especially tactics such as reverse social engineering emerges as a critical insider threat vector that traditional anomaly detection frameworks are ill-equipped to address. This necessitates a deeper understanding of how various social engineering techniques operate internally, manipulating trust and organizational dynamics, thereby justifying the need for adaptive, context-aware detection models capable of capturing socio-cognitive manipulation patterns. 4.3.2 Typologies and Insider Threat Exploitation Previous research in insider threat detection has long relied on typologies to classify threat actors based on factors such as intent, access rights, and behavioural patterns. Notable works such as those by (Zhang et al., 2021 ; andYuan & Wu, 2021 ) typically distinguish between malicious insiders, who act with deliberate intent to cause harm; unintentional insiders, who are driven by negligence or error; masqueraders, who impersonate legitimate users to gain access; and traitorous insiders, who abuse legitimate privileges to breach trust(Maestre Vidal & Sotelo Monge, 2020 ). These typologies have provided foundational models for deep learning-based anomaly detection systems. However, a critical oversight in these frameworks is the exclusion of socially engineered insiders specifically those manipulated from within through psychological and procedural exploitation. Social engineering is often studied as an external threat vector, yet internal adaptation of these techniques such as reverse social engineering represents an increasingly sophisticated attack surface where malicious insiders manipulate compliant colleagues to perform harmful actions unknowingly. This conceptual gap undermines the efficacy of current detection models, which are typically trained to flag technical anomalies, not intent distortion or trust exploitation. While (Kamatchi and Uma, 2025b ) emphasize the integration of psychological profiling in DL models, and (Alzaabi and Mehmood, 2024b ; Song et al., 2024) explore hybrid models for behavioural sequence detection, none of these studies directly address manipulation-based behavioural dynamics. This study builds upon and extends existing typological research by formally introducing socially engineered insiders into the insider threat landscape and highlighting the need for deep learning systems capable of modelling not just anomalies, but behavioural intent, psychological influence, and relational context. The typologies and techniques analysed in this section illustrate that traditional, technically focused insider threat detection systems are inadequate for addressing threats rooted in human vulnerabilities, emotional responses, and social manipulation. Internal adaptations of phishing, pretexting, baiting, quid pro quo, and particularly reverse social engineering, reveal how malicious insiders can weaponize trust, hierarchy, and routine processes to induce harm without triggering conventional anomaly thresholds. Therefore, detecting such threats demands a paradigm shift from anomaly only models to contextual deep learning models that integrate: Communication and linguistic behaviour analysis, Sentiment and emotional state tracking, Procedural and policy deviation detection, and Relational graph modelling across organizational hierarchies. These insights substantiate the novel contribution of this study and reinforce the need for the context-aware, multi-signal deep learning architecture presented in Section 5 , capable of identifying not just behavioural anomalies, but the sociocognitive signatures of internal social engineering threats. 4.4. Deep Learning Models and Architectural Trends The evolution of deep learning (DL) architectures in insider threat detection reflects a progressive effort to model increasingly complex behavioural and relational patterns. Early insider threat detection models primarily employed sequential deep learning techniques such as Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs) to model temporal dependencies within system usage patterns and behavioural logs (Gayathri et al., 2024a ;Randive et al., 2023). These models were effective in detecting basic deviations in access times, file manipulation sequences, and login behaviours. Subsequent advancements introduced attention-based mechanisms to enhance sequential models, allowing detection systems to dynamically focus on critical segments of behavioural sequences(Song et al., 2024); (Amiri-Zarandi et al., 2023b ). LSTM-Attention hybrids and Transformer-based frameworks emerged to better handle long-range dependencies and highlight subtle, critical behaviour deviations often overlooked in standard LSTM architectures. Graph Neural Networks (GNNs) marked another significant shift by enabling the modelling of relational structures among users, devices, and files. Studies such as DTITD (Wang and El Saddik, 2023; He et al., 2024b ) utilized GNNs to capture the graph-like interactions within enterprise systems, facilitating the detection of collective behavioural anomalies and relational misuse patterns. Emerging approaches have further explored meta-learning and adversarial frameworks. GMFITD (Li et al., 2024b ) applied graph meta-learning to address few-shot learning challenges in insider threat scenarios where labelled malicious samples are scarce. Similarly, adversarial autoencoder models like AUTH (Zhu et al., 2024 ) attempted to enhance generalization by learning robust latent representations across different insider behaviour domains. However, despite these architectural innovations, several critical limitations persist: Technical Anchoring : Most models, regardless of architectural sophistication, remain focused on syntactic anomalies in access or system usage data, failing to model the semantic or psychological context of insider actions (Liu et al., 2019 b). Interpretability Challenges : As models grow deeper and more complex (e.g., Transformers with multi-head attention layers, hybrid GNNs), their interpretability diminishes, making it difficult for security analysts to trust or validate their outputs (Chattopadhyay et al., 2024). Neglect of Human Factors : None of the reviewed DL architectures explicitly incorporate psychological signals such as emotional stress, trust dynamics, or social manipulation indicators, which are crucial for detecting reverse social engineering or manipulated insiders. While architectural sophistication has improved model performance on technical benchmarks, the critical dimension of why an insider’s behaviour deviates especially under social influence remains largely unmodeled. Models can detect that an anomaly occurred but cannot explain whether it was induced through cognitive manipulation, emotional coercion, or hierarchical trust exploitation. The evolution from LSTM-based sequential models to GNNs, Transformers, and Meta-Learning has advanced the field of insider threat detection in modelling complexity and relational reasoning. However, the absence of sociotechnical behavioural and organisational relational context models across these architectures perpetuates a blind spot toward insider threats induced through internal social engineering. 4.5. Dataset Challenges and Real-world Challenges Dataset availability and real-world challenges represent one of the most persistent and critical challenges in insider threat detection research. The performance and generalizability of deep learning (DL) models are intrinsically tied to the quality, richness, and contextual realism of the datasets used for their training and evaluation. Across the reviewed studies, a heavy dependence on a small number of benchmark datasets was evident, particularly the CERT Insider Threat Dataset (versions 4.2, 5.2, and 6.2), DARPA1999, and, to a lesser extent, the Enron Email Corpus (Soh et al., 2019 ); (Janjua et al., 2021 ). The CERT datasets, although widely used and relatively comprehensive in terms of simulated organizational activities (e.g., email usage, web browsing, file access, USB activities), were synthetically generated in laboratory settings without real-world emotional, relational, or social manipulation dynamics. User behaviours are algorithmically scripted based on predefined "malicious profiles" rather than naturally emerging from complex social interactions or manipulative influences. As a result, while CERT datasets enable the modelling of technical misuse patterns (e.g., data exfiltration, privilege misuse), they are fundamentally inadequate for training models to detect insider threats rooted in cognitive bias, emotional exploitation, or relational trust violations (Kumar, 2024 ). Similarly, the DARPA1999 dataset, originally designed for network intrusion detection, is outdated and lacks the behavioural depth necessary for modern insider threat modelling(Mahmoud et al., 2023 ). The Enron Email Corpus, while offering real-world communication data, suffers from a lack of insider threat labelling and ground truth, making it difficult to systematically study malicious behavioural evolution or social engineering manipulation within organizational contexts. Some studies attempted to address these limitations through data augmentation. For instance, hybrid approaches combining Generative Adversarial Networks (GANs) and synthetic behaviour injection (Song et al., 2024) have been proposed to create more diverse insider activity patterns. However, even these methods largely focus on technical behaviour variation (e.g., varying file access frequencies, login anomalies) rather than simulating complex emotional manipulation or trust exploitation scenarios (R.G. et al., 2024). Furthermore, none of the reviewed datasets explicitly model or annotate internal social engineering events such as trust-based deception, emotional blackmail, or authority pressure-induced access breaches which are central to understanding and detecting reverse social engineering attacks. The overwhelming reliance on technically scripted, behaviourally shallow datasets like CERT and DARPA significantly constrains the development of deep learning models capable of detecting human-cantered insider threats. Without datasets that capture the complexity of emotional, cognitive, and relational manipulation dynamics, models will continue to excel at detecting technical anomalies while remaining blind to the more insidious, socially engineered insider breaches. 4.6. Evaluation Metrics and Model Adaptability Limitations The evaluation of deep learning (DL) models for insider threat detection has traditionally relied on a limited set of standard performance metrics: Precision, Recall (Sensitivity), F1-Score, Accuracy, and occasionally Area Under the Receiver Operating Characteristic Curve (AUC-ROC). Across the reviewed studies, Precision and Recall were consistently used as the primary indicators of model success (Randive and Ramasundaram, 2023 ; (Alzaabi and Mehmood, 2024). While these metrics provide useful information about technical anomaly detection performance such as how well a model can identify unauthorized file access or unusual login times, they fail to capture the nuanced, human-cantered aspects of insider threat behaviour, particularly those induced through social engineering manipulation (C. R. Zhang et al., 2021 ). Several critical limitations were observed: Precision-Recall Trade off Misalignment : Precision and Recall inherently assume that insider threat instances are clearly separable from benign activities based on observable features. However, socially engineered actions, such as a deceived employee inadvertently leaking sensitive data, may exhibit feature patterns almost indistinguishable from normal behaviour. Models evaluated solely on syntactic anomalies risk misclassifying manipulated behaviours as normal (Song et al., 2024; Zhao et al., 2024). Class Imbalance Distortion : Insider threat datasets, such as CERT, exhibit extreme class imbalance, where malicious instances represent a tiny fraction of overall activity. While techniques like SMOTE oversampling or GAN-based data augmentation attempt to mitigate this, Precision, Recall, and F1-Score often fluctuate dramatically with minor changes in threshold settings, leading to misleading assessments of model robustness (Xu et al., 2024). Lack of Contextual Performance Metrics : None of the reviewed studies incorporated context-aware evaluation metrics that consider relational trust breaches, emotional exploitation, or manipulation resilience. Traditional metrics cannot distinguish whether a detected anomaly is due to technical misbehaviour, behavioural drift, or social manipulation, resulting in evaluation blind spots. Adaptability and Temporal Drift : Insider behaviours evolve over time, especially under social engineering influence. However, few models explicitly evaluated adaptability to behavioural drift, cognitive manipulation, or longitudinal changes in trust dynamics. Most models assume stationarity in user behaviour baselines, an assumption invalidated in real-world, socially dynamic environments (Li et al., 2023; Zhang et al., 2024). Emerging evaluation paradigms in cybersecurity anomaly detection, such as explainability measures and trustworthiness metrics such as confidence calibration, causality detection, have not yet been widely adopted in insider threat research, further compounding the gap. Current evaluation metrics while effective for syntactic anomaly classification are inadequate for assessing a model’s ability to detect cognitively manipulated, socially engineered insider threats. Future insider threat detection research must develop and integrate new evaluation frameworks that measure social trust breach detection, emotional manipulation sensitivity, and longitudinal behavioural adaptability. Evaluation Metrics Evaluation metrics play a critical role in benchmarking the effectiveness of insider threat detection models. They provide quantitative evidence of how well a model can identify malicious or manipulated insider behavior while minimizing false alarms. Across the reviewed studies, a wide range of metrics were employed, reflecting the diversity of approaches and the complexity of insider threat detection. The most frequently used metric is accuracy, which measures the overall correctness of classification outcomes. While accuracy provides a straightforward measure, it can be misleading in highly imbalanced datasets, where malicious insider instances are significantly fewer than benign behaviors. To address this imbalance, precision, recall, and the F1-score are often reported in combination. Precision measures the proportion of correctly identified insider threats among all flagged instances, while recall captures the proportion of actual threats that were successfully detected. The F1-score balances these two metrics, offering a more robust assessment in scenarios where both false positives and false negatives are costly. Other metrics have also been applied, albeit less consistently. The Area Under the Curve (AUC) is frequently used to evaluate classification thresholds and discriminate between insider and non-insider activity. More specialized measures, such as Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), appear in studies adopting regression-based approaches. Additional indicators such as detection rate, true positive rate (TPR), false negative rate (FNR), and R-score were reported in fewer cases, reflecting attempts to capture detection effectiveness from multiple perspectives. The overall distribution of these metrics across the reviewed studies is presented in Fig. 5 , which clearly shows the predominance of accuracy and precision, followed by recall, F1-score, and AUC as the most widely used evaluation measures. This indicates that while research has emphasized classification performance, there is less attention to socio-technical measures or human-centric evaluation criteria that could capture the impact of manipulation and behavioral vulnerabilities. 5. Conceptual Architecture for Context-Aware Insider Threat Detection Building on the qualitative synthesis and the critical research gaps identified in Section 4 , this section presents a novel conceptual architecture designed to enhance insider threat detection capabilities through the integration of behavioural, procedural, and psychological signals alongside traditional system event analysis (Kim et al., 2019 ; Yuan & Wu, 2021 ). Recognizing the increasing sophistication of insider attacks particularly those exploiting internal social engineering techniques such as reverse social engineering the proposed framework aims to enable context-aware, adaptive, and explainable deep learning (DL) models (Alzaabi et al., 2023; Al-Shehari et al., 2024 ). The objective is to transcend traditional anomaly detection based solely on technical artifacts and instead move toward human-centric detection that understands emotional manipulation, trust exploitation, and procedural deviations (Kamatchi & Uma, 2025). 5.1. Overview of the Conceptual Architecture The conceptual framework is founded on the principle that insider threats emerge not merely from technical deviations but from complex sociotechnical interactions where human factors, emotional vulnerabilities, and procedural weaknesses are systematically exploited (Pitropakis et al., 2019; Song et al., 2024). As shown in Figure X (conceptual diagram to be inserted later), the framework synthesizes four primary signal streams: Behavioral signals capturing user activity and operational behavior drift (Xu et al., 2024); Procedural signals representing deviations from organizational policies, workflows, and access control norms (Al-Mhiqani et al., 2020); Psychological signals analyzing sentiment, stress levels, and trust dynamics from user communications (Amiri-Zarandi et al., 2023); Traditional system logs providing security event data such as authentication failures, file movements, and network anomalies (Randive et al., 2023). These multi-modal data streams are preprocessed, embedded into structured feature representations, and fused through advanced deep learning models tailored to capture both sequential and relational dependencies, supported by interpretability modules to ensure explainable detection outputs (Zhao et al., 2024). 5.2. Integration of Behavioural, Procedural, and Psychological Signals Behavioural signals are collected from system logs such as login/logout activities, file access patterns, command execution sequences, and web browsing behaviours (Li et al., 2024). These are encoded as temporal sequences to model normal behavioural baselines and detect deviations over time (Song et al., 2024). Procedural signals involve tracking user role assignments, access escalations, deviations from standard operating procedures (SOPs), and policy violations (Pitropakis et al., 2019; Alzaabi et al., 2023). Such procedural anomalies often precede or accompany malicious activity initiated through social engineering manipulation (Kamatchi & Uma, 2025). Psychological signals are extracted using Natural Language Processing (NLP) techniques applied to internal communications (emails, chat records, helpdesk interactions). Sentiment analysis, stress detection, and relationship modelling provide indicators of emotional distress, cognitive bias exploitation, and compromised trust (Amiri-Zarandi et al., 2023; Al-Shehari et al., 2024 ). 5.3. Deep Learning Fusion Model The unified feature representations are processed through a hybrid deep learning architecture composed of: Sequential models (e.g., LSTM, GRU) to capture temporal dependencies in user behaviour and emotional drift (Randive et al., 2023; Xu et al., 2024); Graph Neural Networks (GNNs) to model relational structures between users, resources, and devices, identifying shifts in trust and communication patterns (Ghosh et al., 2023); Attention mechanisms layered on top of sequential models to dynamically prioritize emotionally charged or procedurally anomalous activities (Zhao et al., 2024); Transformer components (optional) to model long-range dependencies across user sessions and organizational workflows (Song et al., 2024). This multi-architecture approach ensures that the system can detect both short-term anomalies (e.g., sudden emotional spike leading to unauthorized access) and long-term behaviour drifts (e.g., gradual erosion of role boundaries) (Alzaabi et al., 2023). 5.4. Explainability and Human-Centric Risk Analysis A major shortcoming of traditional insider threat detection frameworks is their lack of transparency, which undermines trust and forensic usability (Kim et al., 2019 ; Yuan & Wu, 2021 ). The proposed architecture incorporates Explainable AI (XAI) modules such as SHAP (SHapley Additive exPlanations) or LIME (Local Interpretable Model-agnostic Explanations) to attribute anomaly scores to specific features, be they emotional shifts, procedural violations, or technical anomalies (Amiri-Zarandi et al., 2023). Risk profiles generated by the system are not black-box scores but enriched narratives that explain: Whether deviations stem from technical misuse, emotional manipulation, procedural drift, or a combination thereof; How trust relationships or organizational communications have evolved anomalously over time (Ghosh et al., 2023; Alzaabi et al., 2023). This interpretability enhances analyst decision-making, supports incident response processes, and improves organizational resilience against complex insider threats (Kamatchi & Uma, 2025; Al-Shehari et al., 2024 ). 5.5. Summary The proposed conceptual architecture advances insider threat detection beyond traditional technical-centric paradigms (Yuan & Wu, 2021 ). By holistically modeling behavioural, procedural, and psychological dynamics through a context-aware, explainable deep learning framework, it addresses critical gaps identified in the current state of research particularly the detection of insider threats induced through internal social engineering manipulation (Pitropakis et al., 2019; Alzaabi et al., 2023). 6. Summary of Key Findings 6.1. RQ1 : What deep learning models have been applied to insider threat detection, and how have they evolved over time? The review reveals that insider threat detection has primarily employed sequential deep learning models (e.g., LSTM, GRU) to capture temporal patterns in user behaviour logs (Song et al., 2024; Randive et al., 2023). Attention-based enhancements, including Transformer models, improved sensitivity to long-range behavioural dependencies. More recently, Graph Neural Networks (GNNs) and meta-learning approaches have emerged to model relational structures and handle data scarcity (Li et al., 2024; Zhao et al., 2024). However, despite architectural advances, these models overwhelmingly focus on technical anomalies and lack capabilities to detect socio-cognitive manipulation or emotional exploitation critical for internal social engineering threats. 6.2. RQ2 : What datasets are commonly utilized for developing, training, and evaluating these models? The CERT Insider Threat Datasets (v4.2, v5.2, v6.2) are the most widely adopted, followed by DARPA1999 and, to a lesser extent, the Enron Email Corpus. While these datasets support the detection of technical misuse, they are behaviourally shallow, lacking events involving trust exploitation, emotional manipulation, or cognitive deception (Alzaabi et al., 2023; Chattopadhyay et al., 2024). Consequently, models trained on these datasets are inherently biased toward detecting overt technical anomalies and are blind to more subtle, socially engineered insider threats. 6.3. RQ3 : What performance evaluation metrics are employed to assess the effectiveness and robustness of deep learning models? Precision, Recall, F1-Score, and AUC-ROC are the predominant evaluation metrics across studies (Randive et al., 2023; Amiri-Zarandi et al., 2023). While these metrics assess syntactic anomaly detection performance, they do not evaluate a model’s sensitivity to behavioural manipulation, emotional coercion, or relational trust breaches. Furthermore, no reviewed study incorporated context-aware evaluation frameworks that could distinguish between technical misbehaviour and socially manipulated insider actions, highlighting a significant evaluation gap. 6.4. RQ4 : What are the key challenges, limitations, and emerging future research directions in applying deep learning to insider threat detection? Key challenges include: Technical Anchoring: Models are heavily biased toward detecting system-level anomalies, missing human-centric behavioural deviations caused by manipulation. Dataset Realism: The lack of datasets capturing internal social engineering severely restricts the training and evaluation of models for human-factor threats. Metric Misalignment: Existing performance metrics fail to capture sociotechnical behavioural and emotional dynamics in insider behaviour. Adaptability Deficits: Most models assume behavioural stationarity, whereas real-world insiders adapt and evolve, especially under social influence. Emerging directions advocate for the development of adaptive, socio-cognitive deep learning models, integration of sentiment and trust dynamics, and the creation of realistic, socially annotated datasets to bridge the critical gap of detecting insider threats induced by reverse social engineering. 7. Discussion and Conclusion This systematic literature review critically examined deep learning–based insider threat detection techniques, with a particular emphasis on the often-overlooked vector of internal social engineering, especially reverse social engineering. By analyzing 49 peer-reviewed studies published between 2015 and 2025, insider threats were classified into three interconnected dimensions: socio-technical threats, behavioral threats, and human-factor threats. The findings reveal that although deep learning architectures such as LSTM, GRU, Transformers, and Graph Neural Networks have demonstrated substantial capabilities in detecting technical and behavioral anomalies, current approaches largely neglect the subtle cognitive, emotional, and relational manipulation dynamics that underpin internal social engineering attacks. Benchmark datasets such as the CERT Insider Threat Dataset, DARPA1999, and the Enron Email Corpus demonstrate notable limitations in modeling emotional manipulation and procedural deviation scenarios. Likewise, traditional evaluation metrics, including precision, recall, and the F1-score, insufficiently capture the human-centric nuances essential for detecting socially engineered insiders. Collectively, these shortcomings expose a critical conceptual and methodological gap: the absence of adaptive, context-aware, and socio-cognitive deep learning models capable of modeling trust exploitation and emotional manipulation dynamics within organizations. To address these deficiencies, this study proposes a conceptual architecture for insider threat detection that synthesizes behavioral signals, procedural deviations, and cognitive manipulation indicators into a context-aware deep learning framework. By integrating sequential modeling, relational graph analysis, attention mechanisms, and explainable AI components, the proposed framework moves beyond anomaly detection to offer a pathway toward human-centric insider threat detection. Such an approach has the potential to identify not only traditional malicious insiders but also innocent employees manipulated through social engineering into compromising organizational security. Looking forward, several research directions emerge. Future work should focus on the creation of realistic, socially annotated insider threat datasets that capture manipulation-driven behaviors; the development of evaluation metrics sensitive to socio-cognitive attacks; and the design of explainable, multimodal detection systems capable of delivering early warnings of internal social manipulation. As outlined in Table 6 , advancing socio-behavioral deep learning models, adaptive drift detection, and human-centered explainable AI will be critical in bridging the gap between technical and social dimensions of insider threat detection. Moreover, as highlighted in Table 7 , this study distinguishes itself from prior works by foregrounding internal social engineering as a central threat vector, emphasizing enriched datasets, and advocating for hybrid model architectures that integrate sequential, relational, and affective signals. In conclusion, this review provides one of the most comprehensive analyses of insider threats and social engineering to date, introducing a socio-technical taxonomy and a future-oriented research agenda that integrates human vulnerabilities into detection strategies. By moving beyond syntactic anomaly detection toward adaptive, behavioral, and emotional resilience, the findings establish a foundation for next-generation cybersecurity systems that are both technically robust and socio-cognitively informed. Declarations Author Contribution We provide the first systematic literature review explicitly situating social engineering as an internal insider threat vector, thereby expanding the conceptual landscape of insider threat detection.We introduce novel taxonomies of models, datasets, evaluation metrics, and emerging research directions that highlight both the strengths and limitations of deep learning techniques.We propose a conceptual architecture that integrates behavioral, procedural, and psychological signals into adaptive, context-aware, and explainable deep learning models, bridging the gap between purely technical anomaly detection and human-centric vulnerabilities.We believe Machine Learning is an appropriate venue for this work, given its emphasis on advancing foundational and applied methods in machine learning. Our review highlights how machine learning models can be extended beyond technical anomaly detection to socio-technical and human-centric domains, a frontier of increasing importance as AI-driven systems are deployed in sensitive organizational environments. By critically analyzing existing deep learning approaches and proposing a forward-looking, human-aware framework, this article contributes to both the theory and practice of machine learning for security. References Ahmadi-Assalemi, G., Al-Khateeb, H., Epiphaniou, G., & Aggoun, A. (2022). Super Learner Ensemble for Anomaly Detection and Cyber-Risk Quantification in Industrial Control Systems. IEEE Internet of Things Journal , 9 (15), 13279–13297. https://doi.org/10.1109/JIOT.2022.3144127 Alam, A., & Barron, H. (2022). Anomaly Detection Using Deep Learning and Big Data Analytics for the Insider Threat Platform. In K. Arai (Ed.), Lecture Notes in Networks and Systems: Vol. 506 LNNS (Issue Computing Conference on Intelligent Computing, pp. 512–531). https://doi.org/10.1007/978-3-031-10461-9_35 Alharthi, D. N., Hammad, M. M., & Regan, A. C. (2020). A Taxonomy of Social Engineering Defense Mechanisms. In K. Arai, S. Kapoor, & R. Bhatia (Eds.), Advances in Intelligent Systems and Computing: Vol. 1130 AISC (Issue Future of Information and Communication Conference (FICC), pp. 27–41). https://doi.org/10.1007/978-3-030-39442-4_3 Al‐shehari, T., & Alsowail, R. A. (2021). An insider data leakage detection using one‐hot encoding, synthetic minority oversampling and machine learning techniques. Entropy , 23 (10), 1258. https://doi.org/10.3390/e23101258 Al-Shehari, T., Kadrie, M., Al-Mhiqani, M. N., Alfakih, T., Alsalman, H., Uddin, M., Ullah, S. S., & Dandoush, A. (2024). Comparative evaluation of data imbalance addressing techniques for CNN-based insider threat detection. Scientific Reports , 14 (1), 24715. https://doi.org/10.1038/s41598-024-73510-9 Alzaabi, F. R., & Mehmood, A. (2024a). A Review of Recent Advances, Challenges, and Opportunities in Malicious Insider Threat Detection Using Machine Learning Methods. IEEE Access , 12 , 30907–30927. https://doi.org/10.1109/ACCESS.2024.3369906 Alzaabi, F. R., & Mehmood, A. (2024b). A Review of Recent Advances, Challenges, and Opportunities in Malicious Insider Threat Detection Using Machine Learning Methods. IEEE Access , 12 , 30907–30927. https://doi.org/10.1109/ACCESS.2024.3369906 Alzaabi, F. R., & Mehmood, A. (2024c). A Review of Recent Advances, Challenges, and Opportunities in Malicious Insider Threat Detection Using Machine Learning Methods. IEEE Access , 12 , 30907–30927. https://doi.org/10.1109/ACCESS.2024.3369906 Alzu’bi, A., Darwish, O., Albashayreh, A., & Tashtoush, Y. (2025a). Cyberattack event logs classification using deep learning with semantic feature analysis. Computers and Security , 150 (August 2024). https://doi.org/10.1016/j.cose.2024.104222 Alzu’bi, A., Darwish, O., Albashayreh, A., & Tashtoush, Y. (2025b). Cyberattack event logs classification using deep learning with semantic feature analysis. Computers and Security , 150 , 104222. https://doi.org/10.1016/j.cose.2024.104222 Amiri-Zarandi, M., Karimipour, H., & Dara, R. A. (2023a). A federated and explainable approach for insider threat detection in IoT. Internet of Things (Netherlands) , 24 (October), 100965. https://doi.org/10.1016/j.iot.2023.100965 Amiri-Zarandi, M., Karimipour, H., & Dara, R. A. (2023b). A federated and explainable approach for insider threat detection in IoT. Internet of Things (Netherlands) , 24 , 100965. https://doi.org/10.1016/j.iot.2023.100965 Atosha, P. B., Özbilge, E., & Kırsal, Y. (2024). Comparative Analysis of Deep Recurrent Neural Networks for Speech Recognition. 32nd IEEE Conference on Signal Processing and Communications Applications, SIU 2024 - Proceedings , 1–4. https://doi.org/10.1109/SIU61531.2024.10600944 Burrell, D. N., & Nobles, C. (2022). Discovering the Emergence of Technical Sociology in Human Capital Systems and Technology-Driven Organizations. International Journal of Human Capital and Information Technology Professionals , 13 (1). https://doi.org/10.4018/IJHCITP.300324 Cai, X., Wang, Y., Xu, S., Li, H., Zhang, Y., Liu, Z., & Yuan, X. (2024). LAN: Learning Adaptive Neighbors for Real-Time Insider Threat Detection. IEEE Transactions on Information Forensics and Security , 19 , 10157–10172. https://doi.org/10.1109/TIFS.2024.3488527 Cramer, G., Maxam, W. P., & Davis, J. C. (2025). Engineering patterns for Trust and Safety on social media platforms: A case study of Mastodon and Diaspora. Journal of Systems and Software , 222 , 112322. https://doi.org/10.1016/j.jss.2024.112322 Fuertes, W., Arévalo, D., Castro, J. D., Ron, M., Estrada, C. A., Andrade, R., Peña, F. F., & Benavides, E. (2022). Impact of Social Engineering Attacks: A Literature Review. In A. Rocha, C. H. Fajardo-Toro, & J. M. R. Rodriguez (Eds.), Smart Innovation, Systems and Technologies (Vol. 255, Issue Multidisciplinary International Conference of Research Applied to Defense and Security (MICRADS), pp. 25–35). https://doi.org/10.1007/978-981-16-4884-7_3 Gallo, L., Gentile, D., Ruggiero, S., Botta, A., & Ventre, G. (2024). The human factor in phishing: Collecting and analyzing user behavior when reading emails. Computers and Security , 139 . https://doi.org/10.1016/j.cose.2023.103671 Gayathri, R. G., Sajjanhar, A., & Xiang, Y. (2024a). Hybrid deep learning model using SPCAGAN augmentation for insider threat analysis. Expert Systems with Applications , 249 . https://doi.org/10.1016/j.eswa.2024.123533 Gayathri, R. G., Sajjanhar, A., & Xiang, Y. (2024b). Hybrid deep learning model using SPCAGAN augmentation for insider threat analysis. Expert Systems with Applications , 249 . https://doi.org/10.1016/j.eswa.2024.123533 Golda, A., Mekonen, K., Pandey, A., Singh, A., Hassija, V., Chamola, V., & Sikdar, B. (2024). Privacy and Security Concerns in Generative AI: A Comprehensive Survey. IEEE Access , 12 , 48126–48144. https://doi.org/10.1109/ACCESS.2024.3381611 Han, R., Kim, K., Choi, B., & Jeong, Y. (2023). A Study on Detection of Malicious Behavior Based on Host Process Data Using Machine Learning. Applied Sciences (Switzerland) , 13 (7). https://doi.org/10.3390/app13074097 Handri, E. Y., Indra Sensuse, D., & Tarigan, A. (2024). Developing an Agile Cybersecurity Framework With Organizational Culture Approach Using Q Methodology. IEEE Access , 12 , 108835–108850. https://doi.org/10.1109/ACCESS.2024.3432160 He, D., Lv, X., Xu, X., Chan, S., & Choo, K. K. R. (2024a). Double-Layer Detection of Internal Threat in Enterprise Systems Based on Deep Learning. IEEE Transactions on Information Forensics and Security , 19 , 4741–4751. https://doi.org/10.1109/TIFS.2024.3372771 He, D., Lv, X., Xu, X., Chan, S., & Choo, K. K. R. (2024b). Double-Layer Detection of Internal Threat in Enterprise Systems Based on Deep Learning. IEEE Transactions on Information Forensics and Security , 19 , 4741–4751. https://doi.org/10.1109/TIFS.2024.3372771 He, D., Lv, X., Xu, X., Yu, S., Li, D., Chan, S., & Guizani, M. (2022). An Effective Double-Layer Detection System Against Social Engineering Attacks. IEEE Network , 36 (6), 92–98. https://doi.org/10.1109/MNET.105.2100425 Hijji, M., & Alam, G. (2021). A Multivocal Literature Review on Growing Social Engineering Based Cyber-Attacks/Threats during the COVID-19 Pandemic: Challenges and Prospective Solutions. IEEE Access , 9 , 7152–7169. https://doi.org/10.1109/ACCESS.2020.3048839 Hong, W., Yin, J., You, M., Wang, H., Cao, J., Li, J., & Liu, M. (2022). Graph Intelligence Enhanced Bi-Channel Insider Threat Detection. In X. Yuan, G. Bai, C. Alcaraz, & S. Majumdar (Eds.), Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics): Vol. 13787 LNCS (Issues 16th International Conference on Network and System Security (NSS), pp. 86–102). https://doi.org/10.1007/978-3-031-23020-2_5 Hughes-Lartey, K., Li, M., Botchey, F. E., & Qin, Z. (2021). Human factor, a critical weak point in the information security of an organization’s Internet of things. Heliyon , 7 (3), e06522. https://doi.org/10.1016/j.heliyon.2021.e06522 Ivanov, M. A., Kliuchnikova, B. V., Chugunkov, I. V., & Plaksina, A. M. (2021). Phishing Attacks and Protection against Them. In Proceedings of the 2021 IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering, ElConRus 2021 (Issue IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering (ElConRus), pp. 425–428). https://doi.org/10.1109/ElConRus51938.2021.9396693 Janjua, F., Masood, A., Abbas, H., Rashid, I., & Khan, M. M. Z. M. (2021). Textual analysis of traitor-based dataset through semi supervised machine learning. Future Generation Computer Systems , 125 , 652–660. https://doi.org/10.1016/j.future.2021.06.036 Javaheri, D., Fahmideh, M., Chizari, H., Lalbakhsh, P., & Hur, J. (2024). Cybersecurity threats in FinTech: A systematic review. Expert Systems with Applications , 241 (September 2023), 122697. https://doi.org/10.1016/j.eswa.2023.122697 Johnson, D. G., & Verdicchio, M. (2024). The sociotechnical entanglement of AI and values. AI and Society , 40 (1), 67–76. https://doi.org/10.1007/s00146-023-01852-5 Kamatchi, K., & Uma, E. (2025a). Insights into user behavioral-based insider threat detection: systematic review. International Journal of Information Security , 24 (2), 1–35. https://doi.org/10.1007/s10207-025-01002-6 Kamatchi, K., & Uma, E. (2025b). Insights into user behavioral-based insider threat detection: systematic review. International Journal of Information Security , 24 (2), 1–35. https://doi.org/10.1007/s10207-025-01002-6 Kamatchi, K., & Uma, E. (2025c). Securing the edge: privacy-preserving federated learning for insider threats in IoT networks. Journal of Supercomputing , 81 (1). https://doi.org/10.1007/s11227-024-06752-z Kamatchi, K., & Uma, E. (2025d). Securing the edge: privacy-preserving federated learning for insider threats in IoT networks. The Journal of Supercomputing , 81 (1), 246. https://doi.org/10.1007/s11227-024-06752-z Kamruzzaman, A., Thakur, K., Ismat, S., Ali, M. L., Huang, K., & Thakur, H. N. (2023). Social Engineering Incidents and Preventions. In R. Paul (Ed.), 2023 IEEE 13th Annual Computing and Communication Workshop and Conference (CCWC) (Issue IEEE 13th Annual Computing and Communication Workshop and Conference (CCWC), pp. 0494–0498). IEEE. https://doi.org/10.1109/CCWC57344.2023.10099202 Khadka, K., Ullah, A. B., Ma, W., Marroquin, E. M., & Alem, Y. (2023). A Survey on the Principles of Persuasion as a Social Engineering Strategy in Phishing. In J. Hu, G. Min, G. Wang, & N. Georgalas (Eds.), Proceedings - 2023 IEEE 22nd International Conference on Trust, Security and Privacy in Computing and Communications, TrustCom/BigDataSE/CSE/EUC/iSCI 2023 (Issue IEEE 22nd International Conference on Trust, Security and Privacy in Computing and Communications (TrustCom) / BigDataSE Conference / CSE Conference / EUC Conference / ISCI Conference, pp. 1631–1638). https://doi.org/10.1109/TrustCom60117.2023.00222 Kim, B., Ryu, K., & Back, J. (2022). A Generalized Hold Based Countermeasure Against Zero-Dynamics Attack With Application to DC-DC Converter. IEEE Access , 10 , 44923–44933. https://doi.org/10.1109/ACCESS.2022.3168128 Kim, J., Park, M., Kim, H., Cho, S., & Kang, P. (2019). Insider threat detection based on user behavior modeling and anomaly detection algorithms. Applied Sciences (Switzerland) , 9 (19). https://doi.org/10.3390/app9194018 Kim, J. yeol, & Kwon, H. Y. (2022). Threat classification model for security information event management focusing on model efficiency. Computers and Security , 120 , 102789. https://doi.org/10.1016/j.cose.2022.102789 Kitchenham, B. (2007). Kitchenham , B .: Guidelines for performing Systematic Literature Reviews in software engineering . EBSE Technical Report EBSE-2007-01 Guidelines for performing Systematic Literature Reviews in Software Engineering. Icse , January 2007 , 1–57. Koutsouvelis, V., Shiaeles, S., Ghita, B., & Bendiab, G. (2020). Detection of insider threats using artificial intelligence and visualisation. Proceedings of the 2020 IEEE Conference on Network Softwarization: Bridging the Gap Between AI and Network Softwarization, NetSoft 2020 , 437–443. https://doi.org/10.1109/NetSoft48620.2020.9165337 Kumar, A. (2024). Cybersecurity Threat Detection using Machine Learning and Network Analysis. Journal of Artificial Intelligence General Science (JAIGS) ISSN:3006-4023 , 1 (1), 38–46. https://doi.org/10.60087/jaigs.v1i1.p46 Le, D. C., & Nur Zincir-Heywood, A. (2019). Machine learning based insider threat modelling and detection. 2019 IFIP/IEEE Symposium on Integrated Network and Service Management, IM 2019 , 1–6. Le, D. C., & Zincir-Heywood, N. (2021). Exploring anomalous behaviour detection and classification for insider threat identification. International Journal of Network Management , 31 (4). https://doi.org/10.1002/nem.2109 Li, X., Li, L., Li, X., Cai, B., Jia, J., Gao, Y., & Yu, S. (2024a). GMFITD: Graph Meta-Learning for Effective Few-Shot Insider Threat Detection. IEEE Transactions on Information Forensics and Security , 19 , 7161–7175. https://doi.org/10.1109/TIFS.2024.3430106 Li, X., Li, L., Li, X., Cai, B., Jia, J., Gao, Y., & Yu, S. (2024b). GMFITD: Graph Meta-Learning for Effective Few-Shot Insider Threat Detection. IEEE Transactions on Information Forensics and Security , 19 , 7161–7175. https://doi.org/10.1109/TIFS.2024.3430106 Liu, C., Huang, H. Y., Albarracin, D., & Bashir, M. (2019). Who Shares What with Whom? Information Sharing Preferences in the Online and Offline Worlds. In T. Z. Ahram & D. Nicholson (Eds.), Advances in Intelligent Systems and Computing (Vol. 782, Issues 9th International Conference on Applied Human Factors and Ergonomics (AHFE) / International Conference on Human Factors in Cybersecurity, pp. 149–158). https://doi.org/10.1007/978-3-319-94782-2_15 Liu, F., Jiang, X., Wen, Y., Xing, X., Zhang, D., & Meng, D. (2019). Log2vec: A heterogeneous graph embedding based approach for detecting cyber threats within enterprise. Proceedings of the ACM Conference on Computer and Communications Security , 1777–1794. https://doi.org/10.1145/3319535.3363224 Liu, H., Liu, M., Han, L., Sun, H., & Fu, C. (2025). Ripple2Detect: A semantic similarity learning based framework for insider threat multi-step evidence detection. Computers and Security , 154 , 104387. https://doi.org/10.1016/j.cose.2025.104387 Liu, L., Chen, C., Zhang, J., De Vel, O., & Xiang, Y. (2019b). Unsupervised Insider Detection Through Neural Feature Learning and Model Optimisation. In J. K. Liu & X. Huang (Eds.), Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics): Vol. 11928 LNCS (Issues 13th International Conference on Network and System Security (NSS), pp. 18–36). https://doi.org/10.1007/978-3-030-36938-5_2 Maestre Vidal, J., & Sotelo Monge, M. A. (2020). Obfuscation of Malicious Behaviors for Thwarting Masquerade Detection Systems Based on Locality Features. Sensors (Basel, Switzerland) , 20 (7). https://doi.org/10.3390/s20072084 Mahmoud, M., Mannan, M., & Youssef, A. (2023). APTHunter: Detecting Advanced Persistent Threats in Early Stages. Digital Threats: Research and Practice , 4 (1). https://doi.org/10.1145/3559768 Manoharan, P., Hong, W., Yin, J., Wang, H., Zhang, Y., & Ye, W. (2024). Optimising Insider Threat Prediction: Exploring BiLSTM Networks and Sequential Features. Data Science and Engineering , 9 (4), 393–408. https://doi.org/10.1007/s41019-024-00260-z Matsuda, T., Ushigome, R., Sonoda, M., Satoh, H., Hanada, T., Kanahama, N., Eto, M., Ishikawa, H., Ikeda, K., & Katoh, D. (2020). Investigation and user’s web search skill evaluation for eye and mouse movement in phishing of short message. In T. Ahram & W. Karwowski (Eds.), Advances in Intelligent Systems and Computing (Vol. 960, Issues 10th International Conference on Applied Human Factors and Ergonomics (AHFE) / AHFE International Conference on Human Factors in Cybersecurity, pp. 131–136). https://doi.org/10.1007/978-3-030-20488-4_13 Nifakos, S., Chandramouli, K., Nikolaou, C. K., Papachristou, P., Koch, S., Panaousis, E., & Bonacina, S. (2021). Influence of human factors on cyber security within healthcare organisations: A systematic review. Sensors , 21 (15). https://doi.org/10.3390/s21155119 Randive, K. D., & Ramasundaram, M. (2023). MWCapsNet: A novel Multi-level Wavelet Capsule Network for insider threat detection using image representations. Neurocomputing , 553 (July), 126588. https://doi.org/10.1016/j.neucom.2023.126588 Randive, K., Mohan, R., & Sivakrishna, A. M. (2023a). An efficient pattern-based approach for insider threat classification using the image-based feature representation. Journal of Information Security and Applications , 73 , 103434. https://doi.org/10.1016/j.jisa.2023.103434 Randive, K., Mohan, R., & Sivakrishna, A. M. (2023b). An efficient pattern-based approach for insider threat classification using the image-based feature representation. Journal of Information Security and Applications , 73 (January), 103434. https://doi.org/10.1016/j.jisa.2023.103434 Remmide, M. A., Boumahdi, F., & Boustia, N. (2024). Toward a Hybrid Approach Combining Deep Learning and Case-Based Reasoning for Phishing Email Detection. International Journal on Artificial Intelligence Tools , 33 (5). https://doi.org/10.1142/S0218213024500155 S, A., D, S., & G, P. (2023). Malicious insider threat detection using variation of sampling methods for anomaly detection in cloud environment. Computers and Electrical Engineering , 105 (August 2022), 108519. https://doi.org/10.1016/j.compeleceng.2022.108519 Scott, J., & Kyobe, M. (2021). Trends in Cybersecurity Management Issues Related to Human Behaviour and Machine Learning. 2021 International Conference on Electrical, Computer and Energy Technologies (ICECET) , IEEE International Conference on Electrical, Computer, and Energy Technologies (ICECET) , 1–8. https://doi.org/10.1109/ICECET52533.2021.9698626 Sedes, F., & Degrace, J. (2024). Social Engineering and Security: From Human Vulnerabilities to Malicious Threats. International Conference on Wireless and Mobile Computing, Networking and Communications , 301–305. https://doi.org/10.1109/WiMob61911.2024.10770451 Soh, C., Yu, S., Narayanan, A., Duraisamy, S., & Chen, L. (2019). Employee profiling via aspect-based sentiment and network for insider threats detection. Expert Systems with Applications , 135 , 351–361. https://doi.org/10.1016/j.eswa.2019.05.043 Song, S., Gao, N., Zhang, Y., & Ma, C. (2024a). BRITD: behavior rhythm insider threat detection with time awareness and user adaptation. Cybersecurity , 7 (1). https://doi.org/10.1186/s42400-023-00190-9 Song, S., Gao, N., Zhang, Y., & Ma, C. (2024b). BRITD: behavior rhythm insider threat detection with time awareness and user adaptation. Cybersecurity , 7 (1). https://doi.org/10.1186/s42400-023-00190-9 Song, S., Gao, N., Zhang, Y., & Ma, C. (2024c). BRITD: behavior rhythm insider threat detection with time awareness and user adaptation. Cybersecurity , 7 (1). https://doi.org/10.1186/s42400-023-00190-9 Still, J. D., & Cain, A. A. (2020). Over-the-shoulder attack resistant graphical authentication schemes impact on working memory. In T. Ahram & W. Karwowski (Eds.), Advances in Intelligent Systems and Computing (Vol. 960, Issues 10th International Conference on Applied Human Factors and Ergonomics (AHFE) / AHFE International Conference on Human Factors in Cybersecurity, pp. 79–86). https://doi.org/10.1007/978-3-030-20488-4_8 Sui, Y., Wang, X., Zheng, K., Shi, Y., & Cao, S. (2022). Personality Privacy Protection Method of Social Users Based on Generative Adversarial Networks. Computational Intelligence and Neuroscience , 2022 . https://doi.org/10.1155/2022/2419987 Suryotrisongko, H., Musashi, Y., Tsuneda, A., & Sugitani, K. (2022). Robust Botnet DGA Detection: Blending XAI and OSINT for Cyber Threat Intelligence Sharing. IEEE Access , 10 , 34613–34624. https://doi.org/10.1109/ACCESS.2022.3162588 Syafitri, W., Shukur, Z., Mokhtar, U. A., Sulaiman, R., & Ibrahim, M. A. (2022a). Social Engineering Attacks Prevention: A Systematic Literature Review. IEEE Access , 10 , 39325–39343. https://doi.org/10.1109/ACCESS.2022.3162594 Syafitri, W., Shukur, Z., Mokhtar, U. A., Sulaiman, R., & Ibrahim, M. A. (2022b). Social Engineering Attacks Prevention: A Systematic Literature Review. IEEE Access , 10 , 39325–39343. https://doi.org/10.1109/ACCESS.2022.3162594 Taherdoost, H. (2024). Insights into Cybercrime Detection and Response: A Review of Time Factor. Information (Switzerland) , 15 (5), 273. https://doi.org/10.3390/info15050273 Tian, T., Gong, Y., Jiang, B., Liu, J., Feng, H., & Lu, Z. (2023). Insider Threat Detection Based On Heterogeneous Graph Neural Network. Proceedings - 2023 IEEE 22nd International Conference on Trust, Security and Privacy in Computing and Communications, TrustCom/BigDataSE/CSE/EUC/ISCI 2023 , 628–635. https://doi.org/10.1109/TrustCom60117.2023.00096 Tian, T., Zhang, C., Jiang, B., Feng, H., & Lu, Z. (2025). Insider threat detection for specific threat scenarios. Cybersecurity , 8 (1). https://doi.org/10.1186/s42400-024-00321-w Tundis, A., Melnik, M., Naveed, H., & Mühlhäuser, M. (2021). A social media-based over layer on the edge for handling emergency-related events. Computers and Electrical Engineering , 96 , 107570. https://doi.org/10.1016/j.compeleceng.2021.107570 Wang, Z. Q., & El Saddik, A. (2023a). DTITD: An Intelligent Insider Threat Detection Framework Based on Digital Twin and Self-Attention Based Deep Learning Models. IEEE Access , 11 , 114013–114030. https://doi.org/10.1109/ACCESS.2023.3324371 Wang, Z. Q., & El Saddik, A. (2023b). DTITD: An Intelligent Insider Threat Detection Framework Based on Digital Twin and Self-Attention Based Deep Learning Models. IEEE Access , 11 , 114013–114030. https://doi.org/10.1109/ACCESS.2023.3324371 Xiao, H., Zhu, Y., Zhang, B., Lu, Z., Du, D., & Liu, Y. (2024). Unveiling shadows: A comprehensive framework for insider threat detection based on statistical and sequential analysis. Computers and Security , 138 , 103665. https://doi.org/10.1016/j.cose.2023.103665 Yasin, A., Fatima, R., Wen, L., JiangBin, Z., & Niazi, M. (2025). What goes wrong during phishing education? A probe into a game-based assessment with unfavorable results. Entertainment Computing , 52 . https://doi.org/10.1016/j.entcom.2024.100815 Yuan, S., & Wu, X. (2021a). Deep learning for insider threat detection: Review, challenges and opportunities. Computers & Security , 104 , 102221. https://doi.org/10.1016/j.cose.2021.102221 Yuan, S., & Wu, X. (2021). Deep learning for insider threat detection: Review, challenges and opportunities. In Computers and Security (Vol. 104). https://doi.org/10.1016/j.cose.2021.102221 Yuan, S., & Wu, X. (2021c). Deep learning for insider threat detection: Review, challenges and opportunities. In Computers and Security (Vol. 104). https://doi.org/10.1016/j.cose.2021.102221 Zaoui, M., Yousra, B., Yassine, S., Yassine, M., & Karim, O. (2024). A Comprehensive Taxonomy of Social Engineering Attacks and Defense Mechanisms: Toward Effective Mitigation Strategies. IEEE Access , 12 , 72224–72241. https://doi.org/10.1109/ACCESS.2024.3403197 Zhang, C. R., Wang, S., Zhan, D. C., Yu, T. Y., Wang, T. G., & Yin, M. Y. (2021). Detecting Insider Threat from Behavioral Logs Based on Ensemble and Self-Supervised Learning. SECURITY AND COMMUNICATION NETWORKS , 2021 . https://doi.org/10.1155/2021/4148441 WE - Science Citation Index Expanded (SCI-EXPANDED) Zhang, C., Wang, S., Zhan, D., Yu, T., Wang, T., & Yin, M. (2021). Detecting Insider Threat from Behavioral Logs Based on Ensemble and Self-Supervised Learning. Security and Communication Networks , 2021 , 1–11. https://doi.org/10.1155/2021/4148441 Zhu, X., Dong, J., Qi, J., Zhou, Z., Dong, Z., Sun, Y., & Wang, M. (2024). AUTH: An Adversarial Autoencoder Based Unsupervised Insider Threat Detection Scheme for Multisource Logs. IEEE Transactions on Industrial Informatics , 20 (9), 10954–10965. https://doi.org/10.1109/TII.2024.3393491 Additional Declarations No competing interests reported. Supplementary Files Appendix.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7396895","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Systematic Review","associatedPublications":[],"authors":[{"id":506207242,"identity":"300718c7-bce1-473d-816a-16c172060413","order_by":0,"name":"Ishara Barhoson Galadima","email":"data:image/png;base64,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","orcid":"","institution":"Universiti Teknologi Malaysia, Universiti Teknologi Malaysia, Universiti Teknologi Malaysia, Universiti Teknologi","correspondingAuthor":true,"prefix":"","firstName":"Ishara","middleName":"Barhoson","lastName":"Galadima","suffix":""},{"id":506207243,"identity":"97145ab4-bc8b-4b14-9ec2-ab941c69dd4d","order_by":1,"name":"Norafida Bte Ithnin","email":"","orcid":"","institution":"Universiti Teknologi Malaysia, Universiti Teknologi Malaysia, Universiti Teknologi Malaysia, Universiti Teknologi","correspondingAuthor":false,"prefix":"","firstName":"Norafida","middleName":"Bte","lastName":"Ithnin","suffix":""},{"id":506207244,"identity":"96674375-cffa-43b5-b921-ba8a7b60ac88","order_by":2,"name":"Nur Haliza Abdulwahab","email":"","orcid":"","institution":"Universiti Teknologi Malaysia, Universiti Teknologi Malaysia, Universiti Teknologi Malaysia, Universiti Teknologi","correspondingAuthor":false,"prefix":"","firstName":"Nur","middleName":"Haliza","lastName":"Abdulwahab","suffix":""},{"id":506207245,"identity":"f5009e51-a26a-40f9-bcb9-ae507df794cf","order_by":3,"name":"Mohd Zamri Osman","email":"","orcid":"","institution":"Universiti Teknologi Malaysia, Universiti Teknologi Malaysia, Universiti Teknologi Malaysia, Universiti Teknologi","correspondingAuthor":false,"prefix":"","firstName":"Mohd","middleName":"Zamri","lastName":"Osman","suffix":""},{"id":506207246,"identity":"9ec167d7-ea33-412b-a9a6-be045b448069","order_by":4,"name":"Gabi Danlami","email":"","orcid":"","institution":"Universiti Teknologi Malaysia, Universiti Teknologi Malaysia, Universiti Teknologi Malaysia, Universiti Teknologi","correspondingAuthor":false,"prefix":"","firstName":"Gabi","middleName":"","lastName":"Danlami","suffix":""}],"badges":[],"createdAt":"2025-08-18 07:38:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7396895/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7396895/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":90082205,"identity":"b5947660-8a5b-48d6-8abe-4c831da312f5","added_by":"auto","created_at":"2025-08-28 09:14:07","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":25438,"visible":true,"origin":"","legend":"\u003cp\u003eOur PRISMA methodology\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7396895/v1/f07f13cb38ea2b0003f2aef5.png"},{"id":90082513,"identity":"2cb15cfd-7954-4af2-bd6e-5903a9b17dce","added_by":"auto","created_at":"2025-08-28 09:22:07","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":17819,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 1.2 Paper selection steps\u003c/p\u003e","description":"","filename":"1.2.png","url":"https://assets-eu.researchsquare.com/files/rs-7396895/v1/6982459ae66c3d574fa9ec43.png"},{"id":90082211,"identity":"90d20e17-4e4d-409f-9f73-2e0b75d4197c","added_by":"auto","created_at":"2025-08-28 09:14:07","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":13044,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 2: Top Deep Learning Models for Insider Threats detection\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7396895/v1/887e605fc2880e1be2de2fa6.png"},{"id":90082209,"identity":"b7c9df73-ae52-4c0b-b448-6a90fa8d76c5","added_by":"auto","created_at":"2025-08-28 09:14:07","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":22873,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 3. Most used dataset for Insider threat detection\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7396895/v1/73cb6d8724bb39feefe9721f.png"},{"id":90082207,"identity":"b9bc0b95-48e9-47b9-9fda-455863d30f48","added_by":"auto","created_at":"2025-08-28 09:14:07","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":24006,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4. The Number and Percentage of articles included for qualitative synthesis\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7396895/v1/f4d8c3f8d27d3dc0a0e3ee83.png"},{"id":90082210,"identity":"b7c2824a-44a7-450b-a965-e150919f839e","added_by":"auto","created_at":"2025-08-28 09:14:07","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":8426,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 5. Most commonly used evaluation metrics for insider threat evaluation models\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-7396895/v1/3fbbfd747c45d42c42e347f9.png"},{"id":90083567,"identity":"08b3255e-67d8-4469-885d-d6850e882edb","added_by":"auto","created_at":"2025-08-28 09:30:08","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1489295,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7396895/v1/9d61190d-0666-42bf-be1d-268ffb18a7e2.pdf"},{"id":90082512,"identity":"ec01c1be-34cf-4363-af20-28a8d4837707","added_by":"auto","created_at":"2025-08-28 09:22:07","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":29028,"visible":true,"origin":"","legend":"","description":"","filename":"Appendix.docx","url":"https://assets-eu.researchsquare.com/files/rs-7396895/v1/7c86ca258b787ae13aaf9d8a.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep Learning Models and Social Engineering Dynamics in Insider Threat Detection: A Systematic Review","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eThe contemporary digital economy and the rapid evolution of cybersecurity attacks on critical sectors in the global economy ecosystem have profoundly influenced the development of Artificial Intelligence (AI)-driven detection models (Johnson and Verdicchio, 2024). As enterprises embrace digitization, insider threats have become an increasingly critical cybersecurity concern, particularly as attackers shift from purely technical exploits to exploiting human vulnerabilities (Han et al., 2023; Still and Cain, 2020). Human factors in cybersecurity are central to shaping security assurance across all domains. Therefore, the insider threat landscape is now characterized by a challenging dynamic, multifaceted risks factors, where malicious insiders and compromised employees through normal organisational standard business process or procedure can interact, influence and induced behavioural changes with sophisticated technical infrastructures to cause organizational harm (Alzaabi and Mehmood, 2024a). This phenomenon underscores the urgent need to rethink how organizations detect, mitigate, and respond to insider threats in a digital system. The rise of insider threats is further complicated by this increasing sociotechnical integration, where malicious insiders can manipulate innocent employees into performing harmful actions based on organisational behaviour unknowingly (Handri et al., 2024).\u003c/p\u003e\n\u003cp\u003eAlthough traditional insider threat models primarily focus on monitoring privileged access or detecting anomalies, they often overlook the more subtle, socially driven manipulations that weaponize human trust (Cramer et al., 2025; Tundis et al., 2021). As the sophistication of social engineering attacks grows, bypassing technical safeguards by exploiting psychological vulnerabilities, the need for human-factor-centric detection strategies becomes increasingly paramount (Matsuda et al., 2020). However, technological advancement presents a paradox. On one hand, AI models, behavioural analytics, and deep learning techniques equip organizations with powerful detection capabilities. On the other hand, these same advances are leveraged by malicious insiders to orchestrate complex, stealthy, and socially engineered attacks (Kim et al., 2022). This duality has resulted in insider threats that are more persistent, context-aware, and difficult to detect using conventional static models. Emerging technologies such as Generative AI, personalized phishing campaigns, and deepfake communications further exacerbate these risks, illustrating the urgent necessity for dynamic, adaptive defence mechanisms that understand both technical and human behavioural indicators. Despite growing academic and industry interest, existing research on insider threat detection demonstrates critical limitations (Golda et al., 2024). While many studies have advanced deep learning-based detection approaches, they predominantly address technical anomalies and neglect the potential threat of internally driven social engineering attacks (Javaheri et al., 2024; Ivanov et al., 2021). Most frameworks (Cramer et al., 2025; He et al., 2022; Syafitri et al., 2022a; Tundis et al., 2021)\u0026nbsp;view social engineering solely as an external threat vector rather than acknowledging its insidious potential when orchestrated within organizational boundaries. Moreover, research efforts often focus on classifying malicious insider behaviour without adequately incorporating sociotechnical factors such as interpersonal manipulation, emotional exploitation, or behavioural deception (S. Yuan \u0026amp; Wu, 2021a). As a result, insider threats induced by internal social engineering remain critically underexplored in both conceptualization and practical mitigation (Khadka et al., 2023; Remmide et al., 2024; Gallo et al., 2024).\u003c/p\u003e\n\u003cp\u003eConsidering the number of various studies on cybersecurity insider threats, social engineering attacks within organizational boundaries as independent studies, suffices the necessity for a systematic comprehensive analysis of \u0026ldquo;social engineering as an insider threat attack vector\u0026rdquo; to deal with these potent evolving risks\u0026nbsp;(Kamatchi and Uma, 2025a;\u0026nbsp;Hijji \u0026amp; Alam, 2021;\u0026nbsp;Scott and Kyobe, 2021; Nifakos et al., 2021; Syafitri et al., 2022b; Alharthi et al., 2020). To pave this way, our study investigates the emerging class of insider threats induced by internal social engineering and critically examines the effectiveness of deep learning-based models in detecting such threats. While prior studies have predominantly focused on technical anomalies and behavioural deviations, they have paid limited or no attention to the broader sociotechnical dimensions and organizational behavioural factors that can enable malicious insiders to influence non-malicious employees. Specifically, organizational procedural weaknesses such as trust hierarchies, role ambiguities, and policy circumventions can be systematically exploited to facilitate internal manipulation, thus amplifying insider risk. By addressing these overlooked dynamics, this study aims to bridge a critical gap in insider threat detection research, advocating for context-aware, human-centric detection models that capture the subtle socio-behavioural mechanisms underpinning modern insider threats. To the best of our knowledge, this is the most comprehensive systematic literature review that explicitly addresses \u0026ldquo;social engineering as insider threats attack vector\u0026rdquo;, considers threats arising from both technological and human factors, and offers a comparative synthesis of deep learning-based detection approaches targeting these vulnerabilities.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAiming to bridge the gaps identified in previous surveys, this review is guided by the following research questions (\u003cstrong\u003eRQs\u003c/strong\u003e):\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRQ1:\u003c/strong\u003e What deep learning models have been applied to insider threat detection, and how have they evolved over time?\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRQ2:\u003c/strong\u003e What datasets are commonly utilized for developing, training, and evaluating these models?\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eRQ3:\u003c/strong\u003e What performance evaluation metrics are used to assess the effectiveness and robustness of insider threat detection models?\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRQ4:\u003c/strong\u003e What are the key challenges, limitations, and emerging future directions in applying deep learning to insider threat detection?\u003c/p\u003e\n\u003cp\u003eThe main contributions of the paper are summarized as follow:\u003c/p\u003e\n\u003cp\u003e(i)\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Conducting a comprehensive investigation of 49 peer-reviewed articles to identify prevailing deep learning models, benchmark datasets, and evaluation methodologies in insider threat detection.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e(ii)\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Highlighting the significant research gap related to internal social engineering, particularly reverse social engineering, as an overlooked insider threat attack vector.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e(iii)\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Proposing a novel taxonomy of insider threat attack vectors that incorporates both malicious intent and human-factor exploitation.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e(iv)\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Synthesizing the challenges and limitations identified in current deep learning approaches, including issues of class imbalance, model interpretability, and temporal behaviour drift.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e(v)\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Recommending actionable research directions for developing adaptive, human-centric behaviourally enriched, and sociotechnical-aware deep learning models capable of detecting both malicious insiders and manipulated unintentional insiders (innocent employees) aiming to redefine insider threat paradigms and provide a roadmap for advancing detection model capable of mitigating the evolving insider threats landscape.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe remainder of this paper is structured as follows: Section 2 presents a review of related studies. Section 3 outlines the research methodology. Section 4 provides the qualitative synthesis and classification of detection approaches. Section 5 introduces the conceptual architecture for context-aware insider threat detection. Section 6 summarizes the key findings, and Section 7 offers the discussion and conclusion.\u003c/p\u003e"},{"header":"2. Related Studies","content":"\u003cp\u003eStudies in the independent domains of cybersecurity insider threat, social engineering and deep learning for anomaly detection are extensive; however, only a few scholarly works have systematically reviewed deep learning-based frameworks specifically for insider threat detection and even limited number have explored social engineering as an internal threat vector.\u0026nbsp;(Yuan and Wu, 2021a)conducted a comprehensive survey on insider threat detection using deep learning techniques by evaluating the capabilities of deep learning architectures on insider threat detection. Their study primarily focused on how temporal, spatial, and structural behaviour patterns can be captured using deep models. They categorized insider threats into traitors, masqueraders, and unintentional insiders, and highlighted challenges related to dataset realism and model interpretability. However, their taxonomy did not include socially engineered insiders particularly those manipulated through internal deception, such as reverse social engineering. A more behaviour-centric review was presented by\u0026nbsp;(Kamatchi and Uma, 2025a), who investigated multi-modal and user behaviour modeling techniques. They discussed the benefits of incorporating few-shot learning and psychological profiling into deep learning architectures, showing potential for improved detection of nuanced behavioural anomalies. Yet, their study did not examine human-factor exploitation or social manipulation, leaving a gap in understanding behaviour motivated or distorted by trust exploitation (Gayathri et al., 2024a). A systematic review by and Alzaabi \u0026amp; Mehmood, 2024)\u0026nbsp;offered a broad evaluation of malicious insider threat detection through machine learning methods, including hybrid models like DeepTaskAPT that leverage LSTM and CNN fusion. Their analysis emphasized feature fusion and sequence profiling but remained focused on direct malicious intent, rather than scenarios where users act under manipulation or coercion (Gayathri et al., 2024a; S et al., 2023a). Addressing the challenge of class imbalance, (Al-Shehari et al., 2024) evaluated the performance of CNN-based models with oversampling techniques such as SMOTE and ADASYN on insider threat datasets. They demonstrated improved accuracy in classifying malicious behaviours in synthetic data environments, particularly using the CERT dataset. Nevertheless, their model performance was evaluated purely on activity patterns, with no consideration for contextual or\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003epsychological influences behind user behaviour. A Comprehensive Taxonomy of Social Engineering Attacks and Defense Mechanisms was addressed by (Zaoui et al., 2024), who develop comprehensive taxonomies of social engineering attacks and defense mechanisms, with a particular emphasis on identifying effective mitigation strategies to enhance cybersecurity. Although their findings are valuable in mitigating against technical threats, their survey primarily focused on external adversarial vectors, and did not account for internal social manipulation. Human factors were directly examined by (Hughes-Lartey et al., 2021), who confirmed that behavioural vulnerabilities are the most critical point of failure in\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003einformation security, particularly in IoT environments (Kamatchi \u0026amp; Uma, 2025c). Their study validated the statistical relationship between human error and data breaches but stopped short of integrating these findings into AI-driven detection mechanisms missing the opportunity to link behavioural\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003emanipulation to deep learning-based insider threat models.\u0026nbsp;(Alam and Barron, 2022;Koutsouvelis et al., 2020)\u0026nbsp;offered a structured overview of DL methods for insider threat detection, classifying models by input type and detection logic. Although they acknowledged the sophistication of insider behaviours, they did not explore the role of social dynamics, coercion, or trust-based manipulation within organizational hierarchies. Additionally, semantic models such as Log2Vec (F. Liu et al., 2019) and transformer-based log analysis tools (Alzu\u0026rsquo;bi et al., 2025a) showed that pre-trained architectures could successfully model event patterns using embeddings. However, these approaches focus on event semantics, not behavioural motivation, and therefore fail to capture psychologically driven insider activity like that seen in reverse social engineering (Yuan \u0026amp; Wu, 2021). As indicated in Table 1 of our study, none of the existing reviews or detection frameworks comprehensively consider insider threats originating from internal social engineering tactics. While most studies treat social engineering as an external attack, our work positions it as an internal vector, where well-meaning employees are manipulated by adversaries from within the organization. Furthermore, most reviews stop at the technical or behavioural level and do not integrate sociotechnical indicators into their detection method.\u003c/p\u003e\n\u003cp\u003eTo the best of our knowledge, this systematic literature review (SLR) is the first to provide a systematic review of deep learning approaches for insider threat detection with a distinctive focus on threats induced through internal social engineering manipulation. Unlike prior reviews that largely concentrate on malicious insiders or technical anomalies, this study emphasizes the emerging class of induced insider threats to well-intentioned individuals who could be manipulated into harmful actions through techniques such as phishing, pretexting, and reverse social engineering. This review makes four major contributions: (1) a taxonomy of deep learning model architectures applied to insider threat detection, (2) a taxonomy of datasets with attention to behavioural and contextual representation, (3) a taxonomy of evaluation metrics and their limitations, and (4) a forward-looking taxonomy of research directions emphasizing sociotechnical and behavioural model. Additionally, the study proposes a conceptual architecture that integrates behavioural, procedural, and psychological signals with traditional log-based analytics, enabling context-aware and explainable deep learning frameworks. Overall, this work addresses a pressing gap in the field by advocating for adaptive, human-centric detection systems capable of identifying both malicious and socially engineered insider threats in modern organizational environments.\u003c/p\u003e\n\u003cp\u003eTable 1: Comparison between related studies, surveys and this SLR.\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n\u003cp\u003eFig 1.2 representing the paper selection process for our systematic literature review. It shows the proportion of studies retained at each stage from initial identification through final inclusion highlighting the screening rigor applied to ensure only the most relevant 49 studies were included. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFigure 2 presents the distribution of deep learning models employed in insider threat detection studies reviewed in this research. It reveals that Long Short-Term Memory (LSTM) networks are the most commonly used individual model, appearing in approximately 13 studies, which underscores their effectiveness in capturing sequential behavioral patterns in user activity data. Convolutional Neural Networks (CNNs) follow, used in around 7 studies, reflecting their adaptability in processing structured security logs. Emerging models such as Graph Neural Networks (GNNs) and Deep Neural Networks (DNNs) appear in fewer studies, with 4 and 3 mentions respectively, suggesting growing interest in relationship modeling and layered representations. Other models, including Feedforward Neural Networks (FNN), Generative Adversarial Networks (GAN), Autoencoders, Deep Belief Networks (DBN), and Gated Recurrent Units (GRU), show more limited application. Notably, the most frequent category is the combination of LSTM with other models\u0026mdash;appearing in 21 studies\u0026mdash;indicating a strong trend toward hybrid and ensemble architectures aimed at enhancing detection accuracy and context-awareness. This pattern highlights the dominance of sequential modeling approaches while also revealing an underutilization of generative and unsupervised learning methods, especially for detecting subtle, socially engineered insider threats.\u003c/p\u003e"},{"header":"3. Research Methodology","content":"\u003cp\u003eOur Systematic Literature Review (SLR) was conducted using the established guidelines of PRISMA 2020 protocol (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) to ensure rigor, transparency, and reproducibility. The objective was to synthesize and critically analyze deep learning-based approaches for insider threat detection published between January 2015 and April 2025.\u003c/p\u003e\n\u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e illustrates the distribution of datasets used in the studies reviewed for insider threat detection research. A dominant 82% of the studies relied on the CERT Insider Threat Dataset developed by Carnegie Mellon University\u0026rsquo;s Software Engineering Institute. This overwhelming reliance reflects its wide acceptance as a benchmark for insider threat detection, largely due to its structured logs, labeled activities, and availability. However, such heavy dependence also underscores a major limitation in the field, as the dataset lacks realistic social engineering scenarios and context-aware behavioral annotations.\u003c/p\u003e\n\u003cp\u003eOther datasets appear in much smaller proportions. Enterprise Event Logs account for 4% of the studies, followed by Critical Infrastructure Activity Dataset, Simulated Cloud Logs, and Coded Dataset, each contributing 2%. Additional niche datasets such as Sapi Mouse Dataset, Generated Dataset, Augmented Dataset, and the Enron Email Dataset also each represent 2% of the total. This fragmented use of alternative datasets suggests a scarcity of diverse, publicly available insider threat datasets that capture nuanced behavioral patterns or socio-cognitive elements.\u003c/p\u003e\n\u003cp\u003eThe chart highlights a significant research gap: the field lacks representative, socially annotated datasets capable of supporting advanced detection models, especially for insider threats induced by manipulation, trust abuse, and reverse social engineering. Addressing this limitation is critical for building and validating adaptive, context-aware detection frameworks.\u003c/p\u003e\n\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\n \u003ch2\u003e3.1 Review Protocol\u003c/h2\u003e\n \u003cp\u003eThe protocol included inclusion/exclusion criteria, selection strategy, quality assessment, and a systematic approach to data extraction and synthesis using literature review matrix.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\n \u003ch2\u003e3.2 Data Sources and Search Strategy\u003c/h2\u003e\n \u003cp\u003eA comprehensive literature search was conducted across multiple academic databases:\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eIEEE Xplore\u003c/strong\u003e\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eScienceDirect (Elsevier)\u003c/strong\u003e\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eWiley Online Library\u003c/strong\u003e\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eScopus\u003c/strong\u003e\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eWeb of Science\u003c/strong\u003e\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eSpringer\u003c/strong\u003e\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eACM\u003c/strong\u003e\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n \u003cp\u003eTo ensure comprehensive coverage of relevant studies, multiple databases were consulted, including IEEE Xplore, ScienceDirect (Elsevier), Wiley Online Library, Scopus, Web of Science, Springer, and the ACM Digital Library. The distribution of articles retrieved from these sources is presented in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e, which shows that the majority of studies were obtained from ScienceDirect (21%) and IEEE Xplore (17%), followed closely by Wiley Online Library (16%) and Web of Science (14%).\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eSearch String\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e(\u0026quot;insider threat\u0026quot; OR \u0026quot;malicious insider\u0026quot;) AND (detection OR \u0026quot;user behaviour\u0026quot;) AND (\u0026quot;deep learning\u0026quot; OR LSTM OR RNN OR \u0026quot;attention mechanism\u0026quot;)\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\n \u003ch2\u003e3.3 Inclusion and Exclusion Criteria\u003c/h2\u003e\n \u003cp\u003eTo ensure rigor and relevance, clear inclusion and exclusion criteria were applied during the study selection process. Only peer-reviewed, English-language articles published between 2015 and April 2025 were considered, focusing specifically on deep learning models for insider threat detection across enterprise, network, cloud, or IoT environments. Studies were excluded if they addressed unrelated cybersecurity threats such as phishing or spam, were not accessible in full text, or lacked technical implementation and evaluation. The detailed inclusion and exclusion criteria are summarized in Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003einclusion and Exclusion\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eCriteria\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eInclusion\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eExclusion\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePublication Years\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStudies published between 2015 and April 2025\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStudies published before 2015\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLanguage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEnglish only\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNon-English\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFocus Area\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStudies involving deep learning for insider threat detection\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStudies on unrelated cybersecurity threats (e.g., phishing, spam)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAccessibility\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePeer-reviewed, publicly available full-text articles\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAbstract-only or inaccessible documents\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEvaluation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMust include model, dataset, and performance evaluation metrics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eConceptual or opinion pieces without technical implementation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eApplication Context\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStudies focused on enterprise, network, cloud, or IoT insider threat\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStudies focusing solely on external threat detection\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv align=\"left\" class=\"colspec\"\u003e\u003cbr\u003e\u003c/div\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n \u003ch2\u003e3.4 Study Selection Process\u003c/h2\u003e\n \u003cp\u003eThe study selection followed a structured 4-stage PRISMA process:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e\u003cstrong\u003e1. Identification\u003c/strong\u003e: 498 articles were retrieved using the defined search terms from selected databases.\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e\u003cstrong\u003e2. Screening\u003c/strong\u003e: 146 articles were excluded based on title relevance.\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e\u003cstrong\u003e3. Eligibility\u003c/strong\u003e: 69 articles were excluded after abstract screening; 66 were retained for full-text analysis.\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e\u003cstrong\u003e4. Inclusion\u003c/strong\u003e: 20 articles were excluded after full-text review (due to insufficient methodological detail or relevance), resulting in 49 \u003cstrong\u003earticles\u003c/strong\u003e for qualitative synthesis.\u003c/p\u003e\n \u003c/span\u003e\n \u003cp\u003eA PRISMA flow diagram (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e) summarizes the selection process.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003e3.5 Quality Assessment\u003c/h2\u003e\n \u003cp\u003eTo ensure the scientific rigor of included studies, each was evaluated using a 5-point checklist adapted from (Kitchenham, \u003cspan class=\"CitationRef\"\u003e2007\u003c/span\u003e):\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003eClear statement of research objective\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eJustification of the DL technique/model used\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eClarity of dataset description and source\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eReported evaluation methodology and reproducibility\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n \u003ch2\u003e3.6 Data Extraction and Synthesis\u003c/h2\u003e\n \u003cp\u003eA data extraction form was developed to capture the following information:\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003ePublication details (year, authors)\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eDeep learning model(s) used\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eDataset(s) used\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eEvaluation metrics\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eReported challenges, limitations, and proposed solutions\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n \u003cp\u003eDescriptive and thematic analysis techniques were employed to identify patterns, trends, and research gaps. The results are grouped according to the research questions and discussed in Section \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"4. Qualitative Synthesis and Detection Classification","content":"\u003cp\u003eAfter an extensive review of the 49 peer-reviewed studies included in our qualitative synthesis (Step 4), we manually classified insider threat research into three interdependent categories: sociotechnical threats, behavioural threats, and human factor threats. This expanded taxonomy builds upon previous classifications that primarily conceptualized insider threats through technical or behavioural lenses, often overlooking the crucial roles of organizational behaviour and relational dynamics within modern digital enterprises (Kamatchi \u0026amp; Uma, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2025c\u003c/span\u003e). Several studies exemplify the evolving approaches to insider threat detection. (He et al., 2024) proposed a double-layer detection framework combining Long Short-Term Memory (LSTM) and Extreme Gradient Boosting tree (XGBoost) for phishing detection, alongside a Bidirectional LSTM with an Attention mechanism for insider threat detection. Similarly, (Xiao et al., \u003cspan citationid=\"CR81\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) introduced a comprehensive framework integrating statistical and sequential analysis through convolutional attention and a Transformer encoder (CATE) to address prior limitations in capturing temporal relationships. However, these sequential-based approaches still face challenges, including their potential inability to capture the full complexity of user behaviour and their reliance on domain expertise for rule-based statistical models, limiting their adaptability to emerging insider threats. (Song et al., \u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e2024b\u003c/span\u003e) further addressed poor detection performance stemming from ineffective use of behavioural time information by applying deep learning-based autoencoders on behavioural logs. While their approach achieved high prediction accuracy and precision, limitations such as data imbalance, insufficient label availability, and elevated false alarm rates in hybrid methods persist. Building on this gap, and to the best of our knowledge, our systematic review comprehensively integrates sociotechnical, behavioural, and human-factor dimensions into a novel and refined taxonomy for insider threat detection placing particular emphasis on internal social engineering as an emerging, critically underexplored attack vector.\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003e4.1. Sociotechnical Threats\u003c/h2\u003e\u003cp\u003eOur qualitative papers for sociotechnical threats where technical system vulnerabilities are exploited through human interaction represents a foundational layer in contemporary insider threat research. Many studies have concentrated on uncovering anomalies within system logs, access behaviours, privilege escalations, and user-device interactions by employing advanced deep learning models (Le \u0026amp; Nur Zincir-Heywood, \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Early models predominantly relied on sequential analysis, with Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs) proving particularly effective in modeling system usage over time (Yeol Kim \u0026amp; Kwon, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Li et al., 2024; and Manoharan et al., \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). More recently, researchers have integrated Graph Neural Networks (GNNs) to capture complex system relationships (Tian et al., \u003cspan citationid=\"CR77\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). For instance, the Digital Twin Insider Threat Detection framework (Wang and El Saddik, \u003cspan citationid=\"CR79\" class=\"CitationRef\"\u003e2023a\u003c/span\u003e) employed Digital Twins combined with Self-Attention Transformers to model digital footprints of users, identifying anomalous interactions across virtual system replicas. Similarly, the Learning Adaptive Neighbors framework by (Cai et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) and introduction of adaptive neighbour model to detect anomalies within relational system graphs (Hong et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). These approaches demonstrate an increasing sophistication in model not only individual behaviours but also inter-system communication patterns indicative of insider exploitation (Le \u0026amp;Zincir-Heywood, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Despite technical advances, significant gaps remain. The majority of these models treat insider threats as purely technical anomalies, neglecting the sociotechnical reality that many technical breaches are facilitated by human-driven social engineering manipulation (Ahmadi-Assalemi et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). For example, a malicious insider may not overtly escalate privileges but may coerce or deceive a colleague into performing privileged actions on their behalf, actions that mimic legitimate behaviour at a system log level. Current GNN and LSTM models, while capable of detecting privilege misuse or anomalous login patterns, lack the ability to discern the \u003cem\u003esocial context\u003c/em\u003e behind these actions. Moreover, models such as Ripple2Detect (H. Liu et al., \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) attempt to trace multi-step evidence of insider attacks using semantic similarity learning, yet still primarily focus on technical event chaining rather than the underlying human manipulation processes that initiate those events (Sui et al., \u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eAlthough deep learning models for sociotechnical threat detection have evolved in sophistication, from sequential anomaly detection to graph-based relational anomaly learning; they largely remain blind to insider threats initiated through social manipulation. Current architectures fail to incorporate indicators of interrelatedness of social and technical aspects of organisation procedural relational exploitation, or hierarchical trust dynamics, all of which precede many technical anomalies. Bridging this gap requires new models capable of integrating sociotechnical relational behavioural signals with technical activity streams.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e4.2. Behavioural Threats\u003c/h2\u003e\u003cp\u003eBehavioural threat model focuses on identifying deviations from expected user activities, aiming to detect early signs of potential insider risk. In the reviewed literature, behavioural model emerged as a dominant approach for insider threat detection, reflecting a recognition that malicious insiders often betray themselves through subtle, progressive shifts in behaviour rather than sudden technical anomalies. Deep learning models, particularly sequence-based architectures such as Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs), were frequently employed to capture temporal dependencies within user activity logs (Ranadive et al., 2023a). Several frameworks introduced enhancements to basic sequential model to better capture complex behaviour drifts. For instance, (Song et al., \u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e2024a\u003c/span\u003e) proposed a behaviour rhythm detection model incorporating time-awareness and user adaptation, dynamically adjusting anomaly baselines based on evolving user behaviours. Similarly, (Suryotrisongko et al., \u003cspan citationid=\"CR72\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) incorporated multi-step evidence aggregation, connecting temporally dispersed anomalous actions into coherent behavioural chains. Hybrid models combining convolutional layers with sequential architectures also gained popularity. (Atosha et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) employed a CNN-GRU hybrid to simultaneously capture spatial (file movement patterns) and temporal (login session anomalies) features, thereby improving sensitivity to nuanced insider behaviours more holistically (R.G. et al., 2024; Tian et al., \u003cspan citationid=\"CR76\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Despite these advances, critical limitations persist. Firstly, most behavioural models focus heavily on overt anomalies such as unusual login times, off-hours system access, or spikes in data movement. They generally neglect the cognitive and emotional dimensions of behaviour that might precede technical anomalies. For example, an insider gradually coerced or manipulated through reverse social engineering might initially exhibit changes in communication style, responsiveness, or collegial interactions none of which are captured in traditional activity logs. Secondly, while models introduced by (Burrell and Nobles, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) focus on relational aspects by adapting neighbour behaviours dynamically, they still prioritize structural system behaviours rather than human relational dynamics. Most studies also struggle with distinguishing between harmless behaviour drift such as a promotion leading to access to new systems and potential malicious drift induced by internal manipulation.\u003c/p\u003e\u003cp\u003eAlthough behavioural threat detection models increasingly capture sequential and temporal irregularities, they remain technically anchored and blind to behaviour deviations arising from social manipulation, cognitive bias exploitation, or emotional leverage. Addressing this gap requires next-generation models capable of integrating behavioural sentiment analysis, cognitive behavioural profiling, and relational context awareness into temporal anomaly detection models.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e4.3. Insider Human Factors and Social Engineering Challenges\u003c/h2\u003e\u003cp\u003eDespite significant advances in detection technologies, modeling human factors in insider threats especially within social engineering contexts remains a critical challenge. Human vulnerabilities, including cognitive biases, emotional responses, and social dynamics, are often exploited by malicious insiders or inadvertently triggered by unsuspecting colleagues. Yet, these dimensions remain underrepresented in most existing threat detection frameworks. Social engineering attacks, such as phishing and pretexting, are designed to manipulate psychological and social susceptibilities(Taherdoost, \u003cspan citationid=\"CR75\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). However, current insider threat models predominantly focus on technical indicators, neglecting behavioural and relational cues essential to understanding these manipulative tactics(Sedes and Degrace, \u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). As a result, detection models struggle to capture the full complexity of insider risks. Moreover, the unpredictable nature of human behaviour limits the efficacy of static detection techniques, calling for more proactive approaches that anticipate subtle signals of compromise(C. Liu et al., \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Integrating sentiment analysis, communication patterns, and social relational behaviour into deep learning frameworks could significantly enhance the early detection of insider threats. Recognizing that human factors are central to cybersecurity resilience, this study advocates for their systematic inclusion in insider threat detection strategies(Burrell \u0026amp; Nobles, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eHuman factors particularly emotional vulnerability, cognitive biases, and relational trust exploitation remain critically underexplored in deep learning models for insider threat detection. Without modeling how insiders are manipulated internally through social engineering, current detection systems risk overlooking the most subtle yet devastating form of insider attacks: those perpetrated unknowingly by compromised but innocent employees. Bridging this gap demands the development of context-aware, socio-cognitively enriched models capable of detecting not only what an insider does but why they behave differently.\u003c/p\u003e\u003cdiv id=\"Sec14\" class=\"Section3\"\u003e\u003ch2\u003e4.3.1 Social Engineering\u003c/h2\u003e\u003cp\u003eSocial engineering is a multifaceted phenomenon encompassing physical, technical, social, and socio-technical dimensions, rendering it resistant to straightforward detection or mitigation through traditional technical protection measures. In the context of insider threats, this challenge is further amplified, as human operators, technicians, and end-users often the weakest links within organizational information systems become primary targets of manipulation rather than machines, code, or technical systems (Yasin et al., \u003cspan citationid=\"CR82\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). The evolution of digital infrastructures, from monolithic systems to interconnected graphs of loosely coupled microservices, along with the proliferation of social networks and cloud-based applications, has expanded the human attack surface (Amiri-Zarandi et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2023a\u003c/span\u003e). Recent advances in cyber-defense technologies have shifted the attack focus from system vulnerabilities to human vulnerabilities, exploiting trust, emotions, and social relationships within organizations (Fuertes et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Particularly, the widespread adoption of Bring Your Own Device (BYOD) policies and the increased reliance on third-party communication platforms in enterprise environments have broadened the range of access points available for sophisticated social engineering attacks. When these human vulnerabilities are combined with the exploitation of zero-day technical weaknesses, social engineering becomes a formidable attack vector, frequently leveraged by advanced persistent threat actors to compromise even highly secured financial and enterprise systems(Kamruzzaman et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Given these developments, internal social engineering especially tactics such as reverse social engineering emerges as a critical insider threat vector that traditional anomaly detection frameworks are ill-equipped to address. This necessitates a deeper understanding of how various social engineering techniques operate internally, manipulating trust and organizational dynamics, thereby justifying the need for adaptive, context-aware detection models capable of capturing socio-cognitive manipulation patterns.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section3\"\u003e\u003ch2\u003e4.3.2 Typologies and Insider Threat Exploitation\u003c/h2\u003e\u003cp\u003ePrevious research in insider threat detection has long relied on typologies to classify threat actors based on factors such as intent, access rights, and behavioural patterns. Notable works such as those by (Zhang et al., \u003cspan citationid=\"CR87\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; andYuan \u0026amp; Wu, \u003cspan citationid=\"CR84\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) typically distinguish between malicious insiders, who act with deliberate intent to cause harm; unintentional insiders, who are driven by negligence or error; masqueraders, who impersonate legitimate users to gain access; and traitorous insiders, who abuse legitimate privileges to breach trust(Maestre Vidal \u0026amp; Sotelo Monge, \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). These typologies have provided foundational models for deep learning-based anomaly detection systems. However, a critical oversight in these frameworks is the exclusion of socially engineered insiders specifically those manipulated from within through psychological and procedural exploitation. Social engineering is often studied as an external threat vector, yet internal adaptation of these techniques such as reverse social engineering represents an increasingly sophisticated attack surface where malicious insiders manipulate compliant colleagues to perform harmful actions unknowingly. This conceptual gap undermines the efficacy of current detection models, which are typically trained to flag technical anomalies, not intent distortion or trust exploitation. While (Kamatchi and Uma, \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2025b\u003c/span\u003e) emphasize the integration of psychological profiling in DL models, and (Alzaabi and Mehmood, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2024b\u003c/span\u003e; Song et al., 2024) explore hybrid models for behavioural sequence detection, none of these studies directly address manipulation-based behavioural dynamics. This study builds upon and extends existing typological research by formally introducing socially engineered insiders into the insider threat landscape and highlighting the need for deep learning systems capable of modelling not just anomalies, but behavioural intent, psychological influence, and relational context.\u003c/p\u003e\u003cp\u003eThe typologies and techniques analysed in this section illustrate that traditional, technically focused insider threat detection systems are inadequate for addressing threats rooted in human vulnerabilities, emotional responses, and social manipulation. Internal adaptations of phishing, pretexting, baiting, quid pro quo, and particularly reverse social engineering, reveal how malicious insiders can weaponize trust, hierarchy, and routine processes to induce harm without triggering conventional anomaly thresholds. Therefore, detecting such threats demands a paradigm shift from anomaly only models to contextual deep learning models that integrate: Communication and linguistic behaviour analysis, Sentiment and emotional state tracking, Procedural and policy deviation detection, and Relational graph modelling across organizational hierarchies. These insights substantiate the novel contribution of this study and reinforce the need for the context-aware, multi-signal deep learning architecture presented in Section \u003cspan refid=\"Sec19\" class=\"InternalRef\"\u003e5\u003c/span\u003e, capable of identifying not just behavioural anomalies, but the sociocognitive signatures of internal social engineering threats.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e4.4. Deep Learning Models and Architectural Trends\u003c/h2\u003e\u003cp\u003eThe evolution of deep learning (DL) architectures in insider threat detection reflects a progressive effort to model increasingly complex behavioural and relational patterns. Early insider threat detection models primarily employed sequential deep learning techniques such as Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs) to model temporal dependencies within system usage patterns and behavioural logs (Gayathri et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2024a\u003c/span\u003e;Randive et al., 2023). These models were effective in detecting basic deviations in access times, file manipulation sequences, and login behaviours. Subsequent advancements introduced attention-based mechanisms to enhance sequential models, allowing detection systems to dynamically focus on critical segments of behavioural sequences(Song et al., 2024); (Amiri-Zarandi et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2023b\u003c/span\u003e). LSTM-Attention hybrids and Transformer-based frameworks emerged to better handle long-range dependencies and highlight subtle, critical behaviour deviations often overlooked in standard LSTM architectures. Graph Neural Networks (GNNs) marked another significant shift by enabling the modelling of relational structures among users, devices, and files. Studies such as DTITD (Wang and El Saddik, 2023; He et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2024b\u003c/span\u003e) utilized GNNs to capture the graph-like interactions within enterprise systems, facilitating the detection of collective behavioural anomalies and relational misuse patterns. Emerging approaches have further explored meta-learning and adversarial frameworks. GMFITD (Li et al., \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2024b\u003c/span\u003e) applied graph meta-learning to address few-shot learning challenges in insider threat scenarios where labelled malicious samples are scarce. Similarly, adversarial autoencoder models like AUTH (Zhu et al., \u003cspan citationid=\"CR89\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) attempted to enhance generalization by learning robust latent representations across different insider behaviour domains.\u003c/p\u003e\u003cp\u003eHowever, despite these architectural innovations, several critical limitations persist:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eTechnical Anchoring\u003c/b\u003e: Most models, regardless of architectural sophistication, remain focused on syntactic anomalies in access or system usage data, failing to model the semantic or psychological context of insider actions (Liu et al., \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2019\u003c/span\u003eb).\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eInterpretability Challenges\u003c/b\u003e: As models grow deeper and more complex (e.g., Transformers with multi-head attention layers, hybrid GNNs), their interpretability diminishes, making it difficult for security analysts to trust or validate their outputs (Chattopadhyay et al., 2024).\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eNeglect of Human Factors\u003c/b\u003e: None of the reviewed DL architectures explicitly incorporate psychological signals such as emotional stress, trust dynamics, or social manipulation indicators, which are crucial for detecting reverse social engineering or manipulated insiders.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eWhile architectural sophistication has improved model performance on technical benchmarks, the critical dimension of \u003cem\u003ewhy\u003c/em\u003e an insider\u0026rsquo;s behaviour deviates especially under social influence remains largely unmodeled. Models can detect \u003cem\u003ethat\u003c/em\u003e an anomaly occurred but cannot explain \u003cem\u003ewhether\u003c/em\u003e it was induced through cognitive manipulation, emotional coercion, or hierarchical trust exploitation.\u003c/p\u003e\u003cp\u003eThe evolution from LSTM-based sequential models to GNNs, Transformers, and Meta-Learning has advanced the field of insider threat detection in modelling complexity and relational reasoning. However, the absence of sociotechnical behavioural and organisational relational context models across these architectures perpetuates a blind spot toward insider threats induced through internal social engineering.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\u003ch2\u003e4.5. Dataset Challenges and Real-world Challenges\u003c/h2\u003e\u003cp\u003eDataset availability and real-world challenges represent one of the most persistent and critical challenges in insider threat detection research. The performance and generalizability of deep learning (DL) models are intrinsically tied to the quality, richness, and contextual realism of the datasets used for their training and evaluation. Across the reviewed studies, a heavy dependence on a small number of benchmark datasets was evident, particularly the CERT Insider Threat Dataset (versions 4.2, 5.2, and 6.2), DARPA1999, and, to a lesser extent, the Enron Email Corpus (Soh et al., \u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e2019\u003c/span\u003e); (Janjua et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). The CERT datasets, although widely used and relatively comprehensive in terms of simulated organizational activities (e.g., email usage, web browsing, file access, USB activities), were synthetically generated in laboratory settings without real-world emotional, relational, or social manipulation dynamics. User behaviours are algorithmically scripted based on predefined \"malicious profiles\" rather than naturally emerging from complex social interactions or manipulative influences. As a result, while CERT datasets enable the modelling of technical misuse patterns (e.g., data exfiltration, privilege misuse), they are fundamentally inadequate for training models to detect insider threats rooted in cognitive bias, emotional exploitation, or relational trust violations (Kumar, \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Similarly, the DARPA1999 dataset, originally designed for network intrusion detection, is outdated and lacks the behavioural depth necessary for modern insider threat modelling(Mahmoud et al., \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). The Enron Email Corpus, while offering real-world communication data, suffers from a lack of insider threat labelling and ground truth, making it difficult to systematically study malicious behavioural evolution or social engineering manipulation within organizational contexts. Some studies attempted to address these limitations through data augmentation. For instance, hybrid approaches combining Generative Adversarial Networks (GANs) and synthetic behaviour injection (Song et al., 2024) have been proposed to create more diverse insider activity patterns. However, even these methods largely focus on technical behaviour variation (e.g., varying file access frequencies, login anomalies) rather than simulating complex emotional manipulation or trust exploitation scenarios (R.G. et al., 2024). Furthermore, none of the reviewed datasets explicitly model or annotate internal social engineering events such as trust-based deception, emotional blackmail, or authority pressure-induced access breaches which are central to understanding and detecting reverse social engineering attacks.\u003c/p\u003e\u003cp\u003eThe overwhelming reliance on technically scripted, behaviourally shallow datasets like CERT and DARPA significantly constrains the development of deep learning models capable of detecting human-cantered insider threats. Without datasets that capture the complexity of emotional, cognitive, and relational manipulation dynamics, models will continue to excel at detecting technical anomalies while remaining blind to the more insidious, socially engineered insider breaches.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\u003ch2\u003e4.6. Evaluation Metrics and Model Adaptability Limitations\u003c/h2\u003e\u003cp\u003eThe evaluation of deep learning (DL) models for insider threat detection has traditionally relied on a limited set of standard performance metrics: Precision, Recall (Sensitivity), F1-Score, Accuracy, and occasionally Area Under the Receiver Operating Characteristic Curve (AUC-ROC). Across the reviewed studies, Precision and Recall were consistently used as the primary indicators of model success (Randive and Ramasundaram, \u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; (Alzaabi and Mehmood, 2024). While these metrics provide useful information about technical anomaly detection performance such as how well a model can identify unauthorized file access or unusual login times, they fail to capture the nuanced, human-cantered aspects of insider threat behaviour, particularly those induced through social engineering manipulation (C. R. Zhang et al., \u003cspan citationid=\"CR87\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eSeveral critical limitations were observed:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003ePrecision-Recall Trade off Misalignment\u003c/b\u003e: Precision and Recall inherently assume that insider threat instances are clearly separable from benign activities based on observable features. However, socially engineered actions, such as a deceived employee inadvertently leaking sensitive data, may exhibit feature patterns almost indistinguishable from normal behaviour. Models evaluated solely on syntactic anomalies risk misclassifying manipulated behaviours as normal (Song et al., 2024; Zhao et al., 2024).\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eClass Imbalance Distortion\u003c/b\u003e: Insider threat datasets, such as CERT, exhibit extreme class imbalance, where malicious instances represent a tiny fraction of overall activity. While techniques like SMOTE oversampling or GAN-based data augmentation attempt to mitigate this, Precision, Recall, and F1-Score often fluctuate dramatically with minor changes in threshold settings, leading to misleading assessments of model robustness (Xu et al., 2024).\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eLack of Contextual Performance Metrics\u003c/b\u003e: None of the reviewed studies incorporated context-aware evaluation metrics that consider relational trust breaches, emotional exploitation, or manipulation resilience. Traditional metrics cannot distinguish whether a detected anomaly is due to technical misbehaviour, behavioural drift, or social manipulation, resulting in evaluation blind spots.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eAdaptability and Temporal Drift\u003c/b\u003e: Insider behaviours evolve over time, especially under social engineering influence. However, few models explicitly evaluated adaptability to behavioural drift, cognitive manipulation, or longitudinal changes in trust dynamics. Most models assume stationarity in user behaviour baselines, an assumption invalidated in real-world, socially dynamic environments (Li et al., 2023; Zhang et al., 2024).\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eEmerging evaluation paradigms in cybersecurity anomaly detection, such as explainability measures and trustworthiness metrics such as confidence calibration, causality detection, have not yet been widely adopted in insider threat research, further compounding the gap. Current evaluation metrics while effective for syntactic anomaly classification are inadequate for assessing a model\u0026rsquo;s ability to detect cognitively manipulated, socially engineered insider threats. Future insider threat detection research must develop and integrate new evaluation frameworks that measure social trust breach detection, emotional manipulation sensitivity, and longitudinal behavioural adaptability.\u003c/p\u003e\u003cp\u003e\u003cb\u003eEvaluation Metrics\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eEvaluation metrics play a critical role in benchmarking the effectiveness of insider threat detection models. They provide quantitative evidence of how well a model can identify malicious or manipulated insider behavior while minimizing false alarms. Across the reviewed studies, a wide range of metrics were employed, reflecting the diversity of approaches and the complexity of insider threat detection. The most frequently used metric is accuracy, which measures the overall correctness of classification outcomes. While accuracy provides a straightforward measure, it can be misleading in highly imbalanced datasets, where malicious insider instances are significantly fewer than benign behaviors. To address this imbalance, precision, recall, and the F1-score are often reported in combination. Precision measures the proportion of correctly identified insider threats among all flagged instances, while recall captures the proportion of actual threats that were successfully detected. The F1-score balances these two metrics, offering a more robust assessment in scenarios where both false positives and false negatives are costly. Other metrics have also been applied, albeit less consistently. The Area Under the Curve (AUC) is frequently used to evaluate classification thresholds and discriminate between insider and non-insider activity. More specialized measures, such as Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), appear in studies adopting regression-based approaches. Additional indicators such as detection rate, true positive rate (TPR), false negative rate (FNR), and R-score were reported in fewer cases, reflecting attempts to capture detection effectiveness from multiple perspectives. The overall distribution of these metrics across the reviewed studies is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e5\u003c/span\u003e, which clearly shows the predominance of accuracy and precision, followed by recall, F1-score, and AUC as the most widely used evaluation measures. This indicates that while research has emphasized classification performance, there is less attention to socio-technical measures or human-centric evaluation criteria that could capture the impact of manipulation and behavioral vulnerabilities.\u003c/p\u003e\u003c/div\u003e"},{"header":"5. Conceptual Architecture for Context-Aware Insider Threat Detection","content":"\u003cp\u003eBuilding on the qualitative synthesis and the critical research gaps identified in Section \u003cspan refid=\"Sec10\" class=\"InternalRef\"\u003e4\u003c/span\u003e, this section presents a novel conceptual architecture designed to enhance insider threat detection capabilities through the integration of behavioural, procedural, and psychological signals alongside traditional system event analysis (Kim et al., \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Yuan \u0026amp; Wu, \u003cspan citationid=\"CR84\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Recognizing the increasing sophistication of insider attacks particularly those exploiting internal social engineering techniques such as reverse social engineering the proposed framework aims to enable context-aware, adaptive, and explainable deep learning (DL) models (Alzaabi et al., 2023; Al-Shehari et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The objective is to transcend traditional anomaly detection based solely on technical artifacts and instead move toward human-centric detection that understands emotional manipulation, trust exploitation, and procedural deviations (Kamatchi \u0026amp; Uma, 2025).\u003c/p\u003e\u003cdiv id=\"Sec20\" class=\"Section2\"\u003e\u003ch2\u003e5.1. Overview of the Conceptual Architecture\u003c/h2\u003e\u003cp\u003eThe conceptual framework is founded on the principle that insider threats emerge not merely from technical deviations but from complex sociotechnical interactions where human factors, emotional vulnerabilities, and procedural weaknesses are systematically exploited (Pitropakis et al., 2019; Song et al., 2024). As shown in Figure X (conceptual diagram to be inserted later), the framework synthesizes four primary signal streams:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eBehavioral signals\u003c/b\u003e capturing user activity and operational behavior drift (Xu et al., 2024);\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eProcedural signals\u003c/b\u003e representing deviations from organizational policies, workflows, and access control norms (Al-Mhiqani et al., 2020);\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003ePsychological signals\u003c/b\u003e analyzing sentiment, stress levels, and trust dynamics from user communications (Amiri-Zarandi et al., 2023);\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eTraditional system logs\u003c/b\u003e providing security event data such as authentication failures, file movements, and network anomalies (Randive et al., 2023).\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThese multi-modal data streams are preprocessed, embedded into structured feature representations, and fused through advanced deep learning models tailored to capture both sequential and relational dependencies, supported by interpretability modules to ensure explainable detection outputs (Zhao et al., 2024).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e\u003ch2\u003e5.2. Integration of Behavioural, Procedural, and Psychological Signals\u003c/h2\u003e\u003cp\u003eBehavioural signals are collected from system logs such as login/logout activities, file access patterns, command execution sequences, and web browsing behaviours (Li et al., 2024). These are encoded as temporal sequences to model normal behavioural baselines and detect deviations over time (Song et al., 2024).\u003c/p\u003e\u003cp\u003eProcedural signals involve tracking user role assignments, access escalations, deviations from standard operating procedures (SOPs), and policy violations (Pitropakis et al., 2019; Alzaabi et al., 2023). Such procedural anomalies often precede or accompany malicious activity initiated through social engineering manipulation (Kamatchi \u0026amp; Uma, 2025).\u003c/p\u003e\u003cp\u003ePsychological signals are extracted using Natural Language Processing (NLP) techniques applied to internal communications (emails, chat records, helpdesk interactions). Sentiment analysis, stress detection, and relationship modelling provide indicators of emotional distress, cognitive bias exploitation, and compromised trust (Amiri-Zarandi et al., 2023; Al-Shehari et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec22\" class=\"Section2\"\u003e\u003ch2\u003e5.3. Deep Learning Fusion Model\u003c/h2\u003e\u003cp\u003eThe unified feature representations are processed through a hybrid deep learning architecture composed of:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eSequential models\u003c/b\u003e (e.g., LSTM, GRU) to capture temporal dependencies in user behaviour and emotional drift (Randive et al., 2023; Xu et al., 2024);\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eGraph Neural Networks (GNNs)\u003c/b\u003e to model relational structures between users, resources, and devices, identifying shifts in trust and communication patterns (Ghosh et al., 2023);\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eAttention mechanisms\u003c/b\u003e layered on top of sequential models to dynamically prioritize emotionally charged or procedurally anomalous activities (Zhao et al., 2024);\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eTransformer components\u003c/b\u003e (optional) to model long-range dependencies across user sessions and organizational workflows (Song et al., 2024).\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThis multi-architecture approach ensures that the system can detect both short-term anomalies (e.g., sudden emotional spike leading to unauthorized access) and long-term behaviour drifts (e.g., gradual erosion of role boundaries) (Alzaabi et al., 2023).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec23\" class=\"Section2\"\u003e\u003ch2\u003e5.4. Explainability and Human-Centric Risk Analysis\u003c/h2\u003e\u003cp\u003eA major shortcoming of traditional insider threat detection frameworks is their lack of transparency, which undermines trust and forensic usability (Kim et al., \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Yuan \u0026amp; Wu, \u003cspan citationid=\"CR84\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). The proposed architecture incorporates Explainable AI (XAI) modules such as SHAP (SHapley Additive exPlanations) or LIME (Local Interpretable Model-agnostic Explanations) to attribute anomaly scores to specific features, be they emotional shifts, procedural violations, or technical anomalies (Amiri-Zarandi et al., 2023).\u003c/p\u003e\u003cp\u003eRisk profiles generated by the system are not black-box scores but enriched narratives that explain:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eWhether deviations stem from technical misuse, emotional manipulation, procedural drift, or a combination thereof;\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eHow trust relationships or organizational communications have evolved anomalously over time (Ghosh et al., 2023; Alzaabi et al., 2023).\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThis interpretability enhances analyst decision-making, supports incident response processes, and improves organizational resilience against complex insider threats (Kamatchi \u0026amp; Uma, 2025; Al-Shehari et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec24\" class=\"Section2\"\u003e\u003ch2\u003e5.5. Summary\u003c/h2\u003e\u003cp\u003eThe proposed conceptual architecture advances insider threat detection beyond traditional technical-centric paradigms (Yuan \u0026amp; Wu, \u003cspan citationid=\"CR84\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). By holistically modeling behavioural, procedural, and psychological dynamics through a context-aware, explainable deep learning framework, it addresses critical gaps identified in the current state of research particularly the detection of insider threats induced through internal social engineering manipulation (Pitropakis et al., 2019; Alzaabi et al., 2023).\u003c/p\u003e\u003c/div\u003e"},{"header":"6. Summary of Key Findings","content":"\u003cp\u003e\u003cb\u003e6.1. RQ1\u003c/b\u003e: What deep learning models have been applied to insider threat detection, and how have they evolved over time?\u003c/p\u003e\u003cp\u003eThe review reveals that insider threat detection has primarily employed sequential deep learning models (e.g., LSTM, GRU) to capture temporal patterns in user behaviour logs (Song et al., 2024; Randive et al., 2023). Attention-based enhancements, including Transformer models, improved sensitivity to long-range behavioural dependencies. More recently, Graph Neural Networks (GNNs) and meta-learning approaches have emerged to model relational structures and handle data scarcity (Li et al., 2024; Zhao et al., 2024). However, despite architectural advances, these models overwhelmingly focus on technical anomalies and lack capabilities to detect socio-cognitive manipulation or emotional exploitation critical for internal social engineering threats.\u003c/p\u003e\u003cdiv id=\"Sec26\" class=\"Section2\"\u003e\u003ch2\u003e\u003cb\u003e6.2. RQ2\u003c/b\u003e: What datasets are commonly utilized for developing, training, and evaluating these models?\u003c/h2\u003e\u003cp\u003eThe CERT Insider Threat Datasets (v4.2, v5.2, v6.2) are the most widely adopted, followed by DARPA1999 and, to a lesser extent, the Enron Email Corpus. While these datasets support the detection of technical misuse, they are behaviourally shallow, lacking events involving trust exploitation, emotional manipulation, or cognitive deception (Alzaabi et al., 2023; Chattopadhyay et al., 2024). Consequently, models trained on these datasets are inherently biased toward detecting overt technical anomalies and are blind to more subtle, socially engineered insider threats.\u003c/p\u003e\u003cp\u003e\u003cb\u003e6.3. RQ3\u003c/b\u003e: What performance evaluation metrics are employed to assess the effectiveness and robustness of deep learning models?\u003c/p\u003e\u003cp\u003ePrecision, Recall, F1-Score, and AUC-ROC are the predominant evaluation metrics across studies (Randive et al., 2023; Amiri-Zarandi et al., 2023). While these metrics assess syntactic anomaly detection performance, they do not evaluate a model\u0026rsquo;s sensitivity to behavioural manipulation, emotional coercion, or relational trust breaches. Furthermore, no reviewed study incorporated context-aware evaluation frameworks that could distinguish between technical misbehaviour and socially manipulated insider actions, highlighting a significant evaluation gap.\u003c/p\u003e\u003cp\u003e\u003cb\u003e6.4. RQ4\u003c/b\u003e: What are the key challenges, limitations, and emerging future research directions in applying deep learning to insider threat detection?\u003c/p\u003e\u003cp\u003eKey challenges include:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eTechnical Anchoring: Models are heavily biased toward detecting system-level anomalies, missing human-centric behavioural deviations caused by manipulation.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eDataset Realism: The lack of datasets capturing internal social engineering severely restricts the training and evaluation of models for human-factor threats.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eMetric Misalignment: Existing performance metrics fail to capture sociotechnical behavioural and emotional dynamics in insider behaviour.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eAdaptability Deficits: Most models assume behavioural stationarity, whereas real-world insiders adapt and evolve, especially under social influence.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eEmerging directions advocate for the development of adaptive, socio-cognitive deep learning models, integration of sentiment and trust dynamics, and the creation of realistic, socially annotated datasets to bridge the critical gap of detecting insider threats induced by reverse social engineering.\u003c/p\u003e\u003c/div\u003e"},{"header":"7. Discussion and Conclusion","content":"\u003cp\u003eThis systematic literature review critically examined deep learning–based insider threat detection techniques, with a particular emphasis on the often-overlooked vector of internal social engineering, especially reverse social engineering. By analyzing 49 peer-reviewed studies published between 2015 and 2025, insider threats were classified into three interconnected dimensions: socio-technical threats, behavioral threats, and human-factor threats. The findings reveal that although deep learning architectures such as LSTM, GRU, Transformers, and Graph Neural Networks have demonstrated substantial capabilities in detecting technical and behavioral anomalies, current approaches largely neglect the subtle cognitive, emotional, and relational manipulation dynamics that underpin internal social engineering attacks. Benchmark datasets such as the CERT Insider Threat Dataset, DARPA1999, and the Enron Email Corpus demonstrate notable limitations in modeling emotional manipulation and procedural deviation scenarios. Likewise, traditional evaluation metrics, including precision, recall, and the F1-score, insufficiently capture the human-centric nuances essential for detecting socially engineered insiders. Collectively, these shortcomings expose a critical conceptual and methodological gap: the absence of adaptive, context-aware, and socio-cognitive deep learning models capable of modeling trust exploitation and emotional manipulation dynamics within organizations.\u003c/p\u003e\n\u003cp\u003eTo address these deficiencies, this study proposes a conceptual architecture for insider threat detection that synthesizes behavioral signals, procedural deviations, and cognitive manipulation indicators into a context-aware deep learning framework. By integrating sequential modeling, relational graph analysis, attention mechanisms, and explainable AI components, the proposed framework moves beyond anomaly detection to offer a pathway toward human-centric insider threat detection. Such an approach has the potential to identify not only traditional malicious insiders but also innocent employees manipulated through social engineering into compromising organizational security.\u003c/p\u003e\n\u003cp\u003eLooking forward, several research directions emerge. Future work should focus on the creation of realistic, socially annotated insider threat datasets that capture manipulation-driven behaviors; the development of evaluation metrics sensitive to socio-cognitive attacks; and the design of explainable, multimodal detection systems capable of delivering early warnings of internal social manipulation. As outlined in \u003cstrong\u003eTable 6\u003c/strong\u003e, advancing socio-behavioral deep learning models, adaptive drift detection, and human-centered explainable AI will be critical in bridging the gap between technical and social dimensions of insider threat detection. Moreover, as highlighted in \u003cstrong\u003eTable 7\u003c/strong\u003e, this study distinguishes itself from prior works by foregrounding internal social engineering as a central threat vector, emphasizing enriched datasets, and advocating for hybrid model architectures that integrate sequential, relational, and affective signals.\u003c/p\u003e\n\u003cp\u003eIn conclusion, this review provides one of the most comprehensive analyses of insider threats and social engineering to date, introducing a socio-technical taxonomy and a future-oriented research agenda that integrates human vulnerabilities into detection strategies. By moving beyond syntactic anomaly detection toward adaptive, behavioral, and emotional resilience, the findings establish a foundation for next-generation cybersecurity systems that are both technically robust and socio-cognitively informed.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eWe provide the first systematic literature review explicitly situating social engineering as an internal insider threat vector, thereby expanding the conceptual landscape of insider threat detection.We introduce novel taxonomies of models, datasets, evaluation metrics, and emerging research directions that highlight both the strengths and limitations of deep learning techniques.We propose a conceptual architecture that integrates behavioral, procedural, and psychological signals into adaptive, context-aware, and explainable deep learning models, bridging the gap between purely technical anomaly detection and human-centric vulnerabilities.We believe Machine Learning is an appropriate venue for this work, given its emphasis on advancing foundational and applied methods in machine learning. Our review highlights how machine learning models can be extended beyond technical anomaly detection to socio-technical and human-centric domains, a frontier of increasing importance as AI-driven systems are deployed in sensitive organizational environments. By critically analyzing existing deep learning approaches and proposing a forward-looking, human-aware framework, this article contributes to both the theory and practice of machine learning for security.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eAhmadi-Assalemi, G., Al-Khateeb, H., Epiphaniou, G., \u0026amp; Aggoun, A. (2022). Super Learner Ensemble for Anomaly Detection and Cyber-Risk Quantification in Industrial Control Systems. \u003cem\u003eIEEE Internet of Things Journal\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(15), 13279\u0026ndash;13297. https://doi.org/10.1109/JIOT.2022.3144127\u003c/li\u003e\n \u003cli\u003eAlam, A., \u0026amp; Barron, H. (2022). Anomaly Detection Using Deep Learning and Big Data Analytics for the Insider Threat Platform. In K. Arai (Ed.), \u003cem\u003eLecture Notes in Networks and Systems: Vol. 506 LNNS\u003c/em\u003e (Issue Computing Conference on Intelligent Computing, pp. 512\u0026ndash;531). https://doi.org/10.1007/978-3-031-10461-9_35\u003c/li\u003e\n \u003cli\u003eAlharthi, D. N., Hammad, M. M., \u0026amp; Regan, A. C. (2020). A Taxonomy of Social Engineering Defense Mechanisms. In K. Arai, S. Kapoor, \u0026amp; R. Bhatia (Eds.), \u003cem\u003eAdvances in Intelligent Systems and Computing: Vol. 1130 AISC\u003c/em\u003e (Issue Future of Information and Communication Conference (FICC), pp. 27\u0026ndash;41). https://doi.org/10.1007/978-3-030-39442-4_3\u003c/li\u003e\n \u003cli\u003eAl‐shehari, T., \u0026amp; Alsowail, R. A. (2021). An insider data leakage detection using one‐hot encoding, synthetic minority oversampling and machine learning techniques. \u003cem\u003eEntropy\u003c/em\u003e, \u003cem\u003e23\u003c/em\u003e(10), 1258. https://doi.org/10.3390/e23101258\u003c/li\u003e\n \u003cli\u003eAl-Shehari, T., Kadrie, M., Al-Mhiqani, M. N., Alfakih, T., Alsalman, H., Uddin, M., Ullah, S. S., \u0026amp; Dandoush, A. (2024). Comparative evaluation of data imbalance addressing techniques for CNN-based insider threat detection. \u003cem\u003eScientific Reports\u003c/em\u003e, \u003cem\u003e14\u003c/em\u003e(1), 24715. https://doi.org/10.1038/s41598-024-73510-9\u003c/li\u003e\n \u003cli\u003eAlzaabi, F. R., \u0026amp; Mehmood, A. (2024a). A Review of Recent Advances, Challenges, and Opportunities in Malicious Insider Threat Detection Using Machine Learning Methods. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e, 30907\u0026ndash;30927. https://doi.org/10.1109/ACCESS.2024.3369906\u003c/li\u003e\n \u003cli\u003eAlzaabi, F. R., \u0026amp; Mehmood, A. (2024b). A Review of Recent Advances, Challenges, and Opportunities in Malicious Insider Threat Detection Using Machine Learning Methods. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e, 30907\u0026ndash;30927. https://doi.org/10.1109/ACCESS.2024.3369906\u003c/li\u003e\n \u003cli\u003eAlzaabi, F. R., \u0026amp; Mehmood, A. (2024c). A Review of Recent Advances, Challenges, and Opportunities in Malicious Insider Threat Detection Using Machine Learning Methods. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e, 30907\u0026ndash;30927. https://doi.org/10.1109/ACCESS.2024.3369906\u003c/li\u003e\n \u003cli\u003eAlzu\u0026rsquo;bi, A., Darwish, O., Albashayreh, A., \u0026amp; Tashtoush, Y. (2025a). Cyberattack event logs classification using deep learning with semantic feature analysis. \u003cem\u003eComputers and Security\u003c/em\u003e, \u003cem\u003e150\u003c/em\u003e(August 2024). https://doi.org/10.1016/j.cose.2024.104222\u003c/li\u003e\n \u003cli\u003eAlzu\u0026rsquo;bi, A., Darwish, O., Albashayreh, A., \u0026amp; Tashtoush, Y. (2025b). Cyberattack event logs classification using deep learning with semantic feature analysis. \u003cem\u003eComputers and Security\u003c/em\u003e, \u003cem\u003e150\u003c/em\u003e, 104222. https://doi.org/10.1016/j.cose.2024.104222\u003c/li\u003e\n \u003cli\u003eAmiri-Zarandi, M., Karimipour, H., \u0026amp; Dara, R. A. (2023a). A federated and explainable approach for insider threat detection in IoT. \u003cem\u003eInternet of Things (Netherlands)\u003c/em\u003e, \u003cem\u003e24\u003c/em\u003e(October), 100965. https://doi.org/10.1016/j.iot.2023.100965\u003c/li\u003e\n \u003cli\u003eAmiri-Zarandi, M., Karimipour, H., \u0026amp; Dara, R. A. (2023b). A federated and explainable approach for insider threat detection in IoT. \u003cem\u003eInternet of Things (Netherlands)\u003c/em\u003e, \u003cem\u003e24\u003c/em\u003e, 100965. https://doi.org/10.1016/j.iot.2023.100965\u003c/li\u003e\n \u003cli\u003eAtosha, P. B., \u0026Ouml;zbilge, E., \u0026amp; Kırsal, Y. (2024). Comparative Analysis of Deep Recurrent Neural Networks for Speech Recognition. \u003cem\u003e32nd IEEE Conference on Signal Processing and Communications Applications, SIU 2024 - Proceedings\u003c/em\u003e, 1\u0026ndash;4. https://doi.org/10.1109/SIU61531.2024.10600944\u003c/li\u003e\n \u003cli\u003eBurrell, D. N., \u0026amp; Nobles, C. (2022). Discovering the Emergence of Technical Sociology in Human Capital Systems and Technology-Driven Organizations. \u003cem\u003eInternational Journal of Human Capital and Information Technology Professionals\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(1). https://doi.org/10.4018/IJHCITP.300324\u003c/li\u003e\n \u003cli\u003eCai, X., Wang, Y., Xu, S., Li, H., Zhang, Y., Liu, Z., \u0026amp; Yuan, X. (2024). LAN: Learning Adaptive Neighbors for Real-Time Insider Threat Detection. \u003cem\u003eIEEE Transactions on Information Forensics and Security\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e, 10157\u0026ndash;10172. https://doi.org/10.1109/TIFS.2024.3488527\u003c/li\u003e\n \u003cli\u003eCramer, G., Maxam, W. P., \u0026amp; Davis, J. C. (2025). Engineering patterns for Trust and Safety on social media platforms: A case study of Mastodon and Diaspora. \u003cem\u003eJournal of Systems and Software\u003c/em\u003e, \u003cem\u003e222\u003c/em\u003e, 112322. https://doi.org/10.1016/j.jss.2024.112322\u003c/li\u003e\n \u003cli\u003eFuertes, W., Ar\u0026eacute;valo, D., Castro, J. D., Ron, M., Estrada, C. A., Andrade, R., Pe\u0026ntilde;a, F. F., \u0026amp; Benavides, E. (2022). Impact of Social Engineering Attacks: A Literature Review. In A. Rocha, C. H. Fajardo-Toro, \u0026amp; J. M. R. Rodriguez (Eds.), \u003cem\u003eSmart Innovation, Systems and Technologies\u003c/em\u003e (Vol. 255, Issue Multidisciplinary International Conference of Research Applied to Defense and Security (MICRADS), pp. 25\u0026ndash;35). https://doi.org/10.1007/978-981-16-4884-7_3\u003c/li\u003e\n \u003cli\u003eGallo, L., Gentile, D., Ruggiero, S., Botta, A., \u0026amp; Ventre, G. (2024). The human factor in phishing: Collecting and analyzing user behavior when reading emails. \u003cem\u003eComputers and Security\u003c/em\u003e, \u003cem\u003e139\u003c/em\u003e. https://doi.org/10.1016/j.cose.2023.103671\u003c/li\u003e\n \u003cli\u003eGayathri, R. G., Sajjanhar, A., \u0026amp; Xiang, Y. (2024a). Hybrid deep learning model using SPCAGAN augmentation for insider threat analysis. \u003cem\u003eExpert Systems with Applications\u003c/em\u003e, \u003cem\u003e249\u003c/em\u003e. https://doi.org/10.1016/j.eswa.2024.123533\u003c/li\u003e\n \u003cli\u003eGayathri, R. G., Sajjanhar, A., \u0026amp; Xiang, Y. (2024b). Hybrid deep learning model using SPCAGAN augmentation for insider threat analysis. \u003cem\u003eExpert Systems with Applications\u003c/em\u003e, \u003cem\u003e249\u003c/em\u003e. https://doi.org/10.1016/j.eswa.2024.123533\u003c/li\u003e\n \u003cli\u003eGolda, A., Mekonen, K., Pandey, A., Singh, A., Hassija, V., Chamola, V., \u0026amp; Sikdar, B. (2024). Privacy and Security Concerns in Generative AI: A Comprehensive Survey. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e, 48126\u0026ndash;48144. https://doi.org/10.1109/ACCESS.2024.3381611\u003c/li\u003e\n \u003cli\u003eHan, R., Kim, K., Choi, B., \u0026amp; Jeong, Y. (2023). A Study on Detection of Malicious Behavior Based on Host Process Data Using Machine Learning. \u003cem\u003eApplied Sciences (Switzerland)\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(7). https://doi.org/10.3390/app13074097\u003c/li\u003e\n \u003cli\u003eHandri, E. Y., Indra Sensuse, D., \u0026amp; Tarigan, A. (2024). Developing an Agile Cybersecurity Framework With Organizational Culture Approach Using Q Methodology. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e, 108835\u0026ndash;108850. https://doi.org/10.1109/ACCESS.2024.3432160\u003c/li\u003e\n \u003cli\u003eHe, D., Lv, X., Xu, X., Chan, S., \u0026amp; Choo, K. K. R. (2024a). Double-Layer Detection of Internal Threat in Enterprise Systems Based on Deep Learning. \u003cem\u003eIEEE Transactions on Information Forensics and Security\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e, 4741\u0026ndash;4751. https://doi.org/10.1109/TIFS.2024.3372771\u003c/li\u003e\n \u003cli\u003eHe, D., Lv, X., Xu, X., Chan, S., \u0026amp; Choo, K. K. R. (2024b). Double-Layer Detection of Internal Threat in Enterprise Systems Based on Deep Learning. \u003cem\u003eIEEE Transactions on Information Forensics and Security\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e, 4741\u0026ndash;4751. https://doi.org/10.1109/TIFS.2024.3372771\u003c/li\u003e\n \u003cli\u003eHe, D., Lv, X., Xu, X., Yu, S., Li, D., Chan, S., \u0026amp; Guizani, M. (2022). An Effective Double-Layer Detection System Against Social Engineering Attacks. \u003cem\u003eIEEE Network\u003c/em\u003e, \u003cem\u003e36\u003c/em\u003e(6), 92\u0026ndash;98. https://doi.org/10.1109/MNET.105.2100425\u003c/li\u003e\n \u003cli\u003eHijji, M., \u0026amp; Alam, G. (2021). A Multivocal Literature Review on Growing Social Engineering Based Cyber-Attacks/Threats during the COVID-19 Pandemic: Challenges and Prospective Solutions. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e, 7152\u0026ndash;7169. https://doi.org/10.1109/ACCESS.2020.3048839\u003c/li\u003e\n \u003cli\u003eHong, W., Yin, J., You, M., Wang, H., Cao, J., Li, J., \u0026amp; Liu, M. (2022). Graph Intelligence Enhanced Bi-Channel Insider Threat Detection. In X. Yuan, G. Bai, C. Alcaraz, \u0026amp; S. Majumdar (Eds.), \u003cem\u003eLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics): Vol. 13787 LNCS\u003c/em\u003e (Issues 16th International Conference on Network and System Security (NSS), pp. 86\u0026ndash;102). https://doi.org/10.1007/978-3-031-23020-2_5\u003c/li\u003e\n \u003cli\u003eHughes-Lartey, K., Li, M., Botchey, F. E., \u0026amp; Qin, Z. (2021). Human factor, a critical weak point in the information security of an organization\u0026rsquo;s Internet of things. \u003cem\u003eHeliyon\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e(3), e06522. https://doi.org/10.1016/j.heliyon.2021.e06522\u003c/li\u003e\n \u003cli\u003eIvanov, M. A., Kliuchnikova, B. V., Chugunkov, I. V., \u0026amp; Plaksina, A. M. (2021). Phishing Attacks and Protection against Them. In \u003cem\u003eProceedings of the 2021 IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering, ElConRus 2021\u003c/em\u003e (Issue IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering (ElConRus), pp. 425\u0026ndash;428). https://doi.org/10.1109/ElConRus51938.2021.9396693\u003c/li\u003e\n \u003cli\u003eJanjua, F., Masood, A., Abbas, H., Rashid, I., \u0026amp; Khan, M. M. Z. M. (2021). Textual analysis of traitor-based dataset through semi supervised machine learning. \u003cem\u003eFuture Generation Computer Systems\u003c/em\u003e, \u003cem\u003e125\u003c/em\u003e, 652\u0026ndash;660. https://doi.org/10.1016/j.future.2021.06.036\u003c/li\u003e\n \u003cli\u003eJavaheri, D., Fahmideh, M., Chizari, H., Lalbakhsh, P., \u0026amp; Hur, J. (2024). Cybersecurity threats in FinTech: A systematic review. \u003cem\u003eExpert Systems with Applications\u003c/em\u003e, \u003cem\u003e241\u003c/em\u003e(September 2023), 122697. https://doi.org/10.1016/j.eswa.2023.122697\u003c/li\u003e\n \u003cli\u003eJohnson, D. G., \u0026amp; Verdicchio, M. (2024). The sociotechnical entanglement of AI and values. \u003cem\u003eAI and Society\u003c/em\u003e, \u003cem\u003e40\u003c/em\u003e(1), 67\u0026ndash;76. https://doi.org/10.1007/s00146-023-01852-5\u003c/li\u003e\n \u003cli\u003eKamatchi, K., \u0026amp; Uma, E. (2025a). Insights into user behavioral-based insider threat detection: systematic review. \u003cem\u003eInternational Journal of Information Security\u003c/em\u003e, \u003cem\u003e24\u003c/em\u003e(2), 1\u0026ndash;35. https://doi.org/10.1007/s10207-025-01002-6\u003c/li\u003e\n \u003cli\u003eKamatchi, K., \u0026amp; Uma, E. (2025b). Insights into user behavioral-based insider threat detection: systematic review. \u003cem\u003eInternational Journal of Information Security\u003c/em\u003e, \u003cem\u003e24\u003c/em\u003e(2), 1\u0026ndash;35. https://doi.org/10.1007/s10207-025-01002-6\u003c/li\u003e\n \u003cli\u003eKamatchi, K., \u0026amp; Uma, E. (2025c). Securing the edge: privacy-preserving federated learning for insider threats in IoT networks. \u003cem\u003eJournal of Supercomputing\u003c/em\u003e, \u003cem\u003e81\u003c/em\u003e(1). https://doi.org/10.1007/s11227-024-06752-z\u003c/li\u003e\n \u003cli\u003eKamatchi, K., \u0026amp; Uma, E. (2025d). Securing the edge: privacy-preserving federated learning for insider threats in IoT networks. \u003cem\u003eThe Journal of Supercomputing\u003c/em\u003e, \u003cem\u003e81\u003c/em\u003e(1), 246. https://doi.org/10.1007/s11227-024-06752-z\u003c/li\u003e\n \u003cli\u003eKamruzzaman, A., Thakur, K., Ismat, S., Ali, M. L., Huang, K., \u0026amp; Thakur, H. N. (2023). Social Engineering Incidents and Preventions. In R. Paul (Ed.), \u003cem\u003e2023 IEEE 13th Annual Computing and Communication Workshop and Conference (CCWC)\u003c/em\u003e (Issue IEEE 13th Annual Computing and Communication Workshop and Conference (CCWC), pp. 0494\u0026ndash;0498). IEEE. https://doi.org/10.1109/CCWC57344.2023.10099202\u003c/li\u003e\n \u003cli\u003eKhadka, K., Ullah, A. B., Ma, W., Marroquin, E. M., \u0026amp; Alem, Y. (2023). A Survey on the Principles of Persuasion as a Social Engineering Strategy in Phishing. In J. Hu, G. Min, G. Wang, \u0026amp; N. Georgalas (Eds.), \u003cem\u003eProceedings - 2023 IEEE 22nd International Conference on Trust, Security and Privacy in Computing and Communications, TrustCom/BigDataSE/CSE/EUC/iSCI 2023\u003c/em\u003e (Issue IEEE 22nd International Conference on Trust, Security and Privacy in Computing and Communications (TrustCom) / BigDataSE Conference / CSE Conference / EUC Conference / ISCI Conference, pp. 1631\u0026ndash;1638). https://doi.org/10.1109/TrustCom60117.2023.00222\u003c/li\u003e\n \u003cli\u003eKim, B., Ryu, K., \u0026amp; Back, J. (2022). A Generalized Hold Based Countermeasure Against Zero-Dynamics Attack With Application to DC-DC Converter. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e, 44923\u0026ndash;44933. https://doi.org/10.1109/ACCESS.2022.3168128\u003c/li\u003e\n \u003cli\u003eKim, J., Park, M., Kim, H., Cho, S., \u0026amp; Kang, P. (2019). Insider threat detection based on user behavior modeling and anomaly detection algorithms. \u003cem\u003eApplied Sciences (Switzerland)\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(19). https://doi.org/10.3390/app9194018\u003c/li\u003e\n \u003cli\u003eKim, J. yeol, \u0026amp; Kwon, H. Y. (2022). Threat classification model for security information event management focusing on model efficiency. \u003cem\u003eComputers and Security\u003c/em\u003e, \u003cem\u003e120\u003c/em\u003e, 102789. https://doi.org/10.1016/j.cose.2022.102789\u003c/li\u003e\n \u003cli\u003eKitchenham, B. (2007). Kitchenham , B .: Guidelines for performing Systematic Literature Reviews in software engineering . EBSE Technical Report EBSE-2007-01 Guidelines for performing Systematic Literature Reviews in Software Engineering. \u003cem\u003eIcse\u003c/em\u003e, \u003cem\u003eJanuary 2007\u003c/em\u003e, 1\u0026ndash;57.\u003c/li\u003e\n \u003cli\u003eKoutsouvelis, V., Shiaeles, S., Ghita, B., \u0026amp; Bendiab, G. (2020). Detection of insider threats using artificial intelligence and visualisation. \u003cem\u003eProceedings of the 2020 IEEE Conference on Network Softwarization: Bridging the Gap Between AI and Network Softwarization, NetSoft 2020\u003c/em\u003e, 437\u0026ndash;443. https://doi.org/10.1109/NetSoft48620.2020.9165337\u003c/li\u003e\n \u003cli\u003eKumar, A. (2024). Cybersecurity Threat Detection using Machine Learning and Network Analysis. \u003cem\u003eJournal of Artificial Intelligence General Science (JAIGS) ISSN:3006-4023\u003c/em\u003e, \u003cem\u003e1\u003c/em\u003e(1), 38\u0026ndash;46. https://doi.org/10.60087/jaigs.v1i1.p46\u003c/li\u003e\n \u003cli\u003eLe, D. C., \u0026amp; Nur Zincir-Heywood, A. (2019). Machine learning based insider threat modelling and detection. \u003cem\u003e2019 IFIP/IEEE Symposium on Integrated Network and Service Management, IM 2019\u003c/em\u003e, 1\u0026ndash;6.\u003c/li\u003e\n \u003cli\u003eLe, D. C., \u0026amp; Zincir-Heywood, N. (2021). Exploring anomalous behaviour detection and classification for insider threat identification. \u003cem\u003eInternational Journal of Network Management\u003c/em\u003e, \u003cem\u003e31\u003c/em\u003e(4). https://doi.org/10.1002/nem.2109\u003c/li\u003e\n \u003cli\u003eLi, X., Li, L., Li, X., Cai, B., Jia, J., Gao, Y., \u0026amp; Yu, S. (2024a). GMFITD: Graph Meta-Learning for Effective Few-Shot Insider Threat Detection. \u003cem\u003eIEEE Transactions on Information Forensics and Security\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e, 7161\u0026ndash;7175. https://doi.org/10.1109/TIFS.2024.3430106\u003c/li\u003e\n \u003cli\u003eLi, X., Li, L., Li, X., Cai, B., Jia, J., Gao, Y., \u0026amp; Yu, S. (2024b). GMFITD: Graph Meta-Learning for Effective Few-Shot Insider Threat Detection. \u003cem\u003eIEEE Transactions on Information Forensics and Security\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e, 7161\u0026ndash;7175. https://doi.org/10.1109/TIFS.2024.3430106\u003c/li\u003e\n \u003cli\u003eLiu, C., Huang, H. Y., Albarracin, D., \u0026amp; Bashir, M. (2019). Who Shares What with Whom? Information Sharing Preferences in the Online and Offline Worlds. In T. Z. Ahram \u0026amp; D. Nicholson (Eds.), \u003cem\u003eAdvances in Intelligent Systems and Computing\u003c/em\u003e (Vol. 782, Issues 9th International Conference on Applied Human Factors and Ergonomics (AHFE) / International Conference on Human Factors in Cybersecurity, pp. 149\u0026ndash;158). https://doi.org/10.1007/978-3-319-94782-2_15\u003c/li\u003e\n \u003cli\u003eLiu, F., Jiang, X., Wen, Y., Xing, X., Zhang, D., \u0026amp; Meng, D. (2019). Log2vec: A heterogeneous graph embedding based approach for detecting cyber threats within enterprise. \u003cem\u003eProceedings of the ACM Conference on Computer and Communications Security\u003c/em\u003e, 1777\u0026ndash;1794. https://doi.org/10.1145/3319535.3363224\u003c/li\u003e\n \u003cli\u003eLiu, H., Liu, M., Han, L., Sun, H., \u0026amp; Fu, C. (2025). Ripple2Detect: A semantic similarity learning based framework for insider threat multi-step evidence detection. \u003cem\u003eComputers and Security\u003c/em\u003e, \u003cem\u003e154\u003c/em\u003e, 104387. https://doi.org/10.1016/j.cose.2025.104387\u003c/li\u003e\n \u003cli\u003eLiu, L., Chen, C., Zhang, J., De Vel, O., \u0026amp; Xiang, Y. (2019b). Unsupervised Insider Detection Through Neural Feature Learning and Model Optimisation. In J. K. Liu \u0026amp; X. Huang (Eds.), \u003cem\u003eLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics): Vol. 11928 LNCS\u003c/em\u003e (Issues 13th International Conference on Network and System Security (NSS), pp. 18\u0026ndash;36). https://doi.org/10.1007/978-3-030-36938-5_2\u003c/li\u003e\n \u003cli\u003eMaestre Vidal, J., \u0026amp; Sotelo Monge, M. A. (2020). Obfuscation of Malicious Behaviors for Thwarting Masquerade Detection Systems Based on Locality Features. \u003cem\u003eSensors (Basel, Switzerland)\u003c/em\u003e, \u003cem\u003e20\u003c/em\u003e(7). https://doi.org/10.3390/s20072084\u003c/li\u003e\n \u003cli\u003eMahmoud, M., Mannan, M., \u0026amp; Youssef, A. (2023). APTHunter: Detecting Advanced Persistent Threats in Early Stages. \u003cem\u003eDigital Threats: Research and Practice\u003c/em\u003e, \u003cem\u003e4\u003c/em\u003e(1). https://doi.org/10.1145/3559768\u003c/li\u003e\n \u003cli\u003eManoharan, P., Hong, W., Yin, J., Wang, H., Zhang, Y., \u0026amp; Ye, W. (2024). Optimising Insider Threat Prediction: Exploring BiLSTM Networks and Sequential Features. \u003cem\u003eData Science and Engineering\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(4), 393\u0026ndash;408. https://doi.org/10.1007/s41019-024-00260-z\u003c/li\u003e\n \u003cli\u003eMatsuda, T., Ushigome, R., Sonoda, M., Satoh, H., Hanada, T., Kanahama, N., Eto, M., Ishikawa, H., Ikeda, K., \u0026amp; Katoh, D. (2020). Investigation and user\u0026rsquo;s web search skill evaluation for eye and mouse movement in phishing of short message. In T. Ahram \u0026amp; W. Karwowski (Eds.), \u003cem\u003eAdvances in Intelligent Systems and Computing\u003c/em\u003e (Vol. 960, Issues 10th International Conference on Applied Human Factors and Ergonomics (AHFE) / AHFE International Conference on Human Factors in Cybersecurity, pp. 131\u0026ndash;136). https://doi.org/10.1007/978-3-030-20488-4_13\u003c/li\u003e\n \u003cli\u003eNifakos, S., Chandramouli, K., Nikolaou, C. K., Papachristou, P., Koch, S., Panaousis, E., \u0026amp; Bonacina, S. (2021). Influence of human factors on cyber security within healthcare organisations: A systematic review. \u003cem\u003eSensors\u003c/em\u003e, \u003cem\u003e21\u003c/em\u003e(15). https://doi.org/10.3390/s21155119\u003c/li\u003e\n \u003cli\u003eRandive, K. D., \u0026amp; Ramasundaram, M. (2023). MWCapsNet: A novel Multi-level Wavelet Capsule Network for insider threat detection using image representations. \u003cem\u003eNeurocomputing\u003c/em\u003e, \u003cem\u003e553\u003c/em\u003e(July), 126588. https://doi.org/10.1016/j.neucom.2023.126588\u003c/li\u003e\n \u003cli\u003eRandive, K., Mohan, R., \u0026amp; Sivakrishna, A. M. (2023a). An efficient pattern-based approach for insider threat classification using the image-based feature representation. \u003cem\u003eJournal of Information Security and Applications\u003c/em\u003e, \u003cem\u003e73\u003c/em\u003e, 103434. https://doi.org/10.1016/j.jisa.2023.103434\u003c/li\u003e\n \u003cli\u003eRandive, K., Mohan, R., \u0026amp; Sivakrishna, A. M. (2023b). An efficient pattern-based approach for insider threat classification using the image-based feature representation. \u003cem\u003eJournal of Information Security and Applications\u003c/em\u003e, \u003cem\u003e73\u003c/em\u003e(January), 103434. https://doi.org/10.1016/j.jisa.2023.103434\u003c/li\u003e\n \u003cli\u003eRemmide, M. A., Boumahdi, F., \u0026amp; Boustia, N. (2024). Toward a Hybrid Approach Combining Deep Learning and Case-Based Reasoning for Phishing Email Detection. \u003cem\u003eInternational Journal on Artificial Intelligence Tools\u003c/em\u003e, \u003cem\u003e33\u003c/em\u003e(5). https://doi.org/10.1142/S0218213024500155\u003c/li\u003e\n \u003cli\u003eS, A., D, S., \u0026amp; G, P. (2023). Malicious insider threat detection using variation of sampling methods for anomaly detection in cloud environment. \u003cem\u003eComputers and Electrical Engineering\u003c/em\u003e, \u003cem\u003e105\u003c/em\u003e(August 2022), 108519. https://doi.org/10.1016/j.compeleceng.2022.108519\u003c/li\u003e\n \u003cli\u003eScott, J., \u0026amp; Kyobe, M. (2021). Trends in Cybersecurity Management Issues Related to Human Behaviour and Machine Learning. \u003cem\u003e2021 International Conference on Electrical, Computer and Energy Technologies (ICECET)\u003c/em\u003e, \u003cem\u003eIEEE International Conference on Electrical, Computer, and Energy Technologies (ICECET)\u003c/em\u003e, 1\u0026ndash;8. https://doi.org/10.1109/ICECET52533.2021.9698626\u003c/li\u003e\n \u003cli\u003eSedes, F., \u0026amp; Degrace, J. (2024). Social Engineering and Security: From Human Vulnerabilities to Malicious Threats. \u003cem\u003eInternational Conference on Wireless and Mobile Computing, Networking and Communications\u003c/em\u003e, 301\u0026ndash;305. https://doi.org/10.1109/WiMob61911.2024.10770451\u003c/li\u003e\n \u003cli\u003eSoh, C., Yu, S., Narayanan, A., Duraisamy, S., \u0026amp; Chen, L. (2019). Employee profiling via aspect-based sentiment and network for insider threats detection. \u003cem\u003eExpert Systems with Applications\u003c/em\u003e, \u003cem\u003e135\u003c/em\u003e, 351\u0026ndash;361. https://doi.org/10.1016/j.eswa.2019.05.043\u003c/li\u003e\n \u003cli\u003eSong, S., Gao, N., Zhang, Y., \u0026amp; Ma, C. (2024a). BRITD: behavior rhythm insider threat detection with time awareness and user adaptation. \u003cem\u003eCybersecurity\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e(1). https://doi.org/10.1186/s42400-023-00190-9\u003c/li\u003e\n \u003cli\u003eSong, S., Gao, N., Zhang, Y., \u0026amp; Ma, C. (2024b). BRITD: behavior rhythm insider threat detection with time awareness and user adaptation. \u003cem\u003eCybersecurity\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e(1). https://doi.org/10.1186/s42400-023-00190-9\u003c/li\u003e\n \u003cli\u003eSong, S., Gao, N., Zhang, Y., \u0026amp; Ma, C. (2024c). BRITD: behavior rhythm insider threat detection with time awareness and user adaptation. \u003cem\u003eCybersecurity\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e(1). https://doi.org/10.1186/s42400-023-00190-9\u003c/li\u003e\n \u003cli\u003eStill, J. D., \u0026amp; Cain, A. A. (2020). Over-the-shoulder attack resistant graphical authentication schemes impact on working memory. In T. Ahram \u0026amp; W. Karwowski (Eds.), \u003cem\u003eAdvances in Intelligent Systems and Computing\u003c/em\u003e (Vol. 960, Issues 10th International Conference on Applied Human Factors and Ergonomics (AHFE) / AHFE International Conference on Human Factors in Cybersecurity, pp. 79\u0026ndash;86). https://doi.org/10.1007/978-3-030-20488-4_8\u003c/li\u003e\n \u003cli\u003eSui, Y., Wang, X., Zheng, K., Shi, Y., \u0026amp; Cao, S. (2022). Personality Privacy Protection Method of Social Users Based on Generative Adversarial Networks. \u003cem\u003eComputational Intelligence and Neuroscience\u003c/em\u003e, \u003cem\u003e2022\u003c/em\u003e. https://doi.org/10.1155/2022/2419987\u003c/li\u003e\n \u003cli\u003eSuryotrisongko, H., Musashi, Y., Tsuneda, A., \u0026amp; Sugitani, K. (2022). Robust Botnet DGA Detection: Blending XAI and OSINT for Cyber Threat Intelligence Sharing. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e, 34613\u0026ndash;34624. https://doi.org/10.1109/ACCESS.2022.3162588\u003c/li\u003e\n \u003cli\u003eSyafitri, W., Shukur, Z., Mokhtar, U. A., Sulaiman, R., \u0026amp; Ibrahim, M. A. (2022a). Social Engineering Attacks Prevention: A Systematic Literature Review. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e, 39325\u0026ndash;39343. https://doi.org/10.1109/ACCESS.2022.3162594\u003c/li\u003e\n \u003cli\u003eSyafitri, W., Shukur, Z., Mokhtar, U. A., Sulaiman, R., \u0026amp; Ibrahim, M. A. (2022b). Social Engineering Attacks Prevention: A Systematic Literature Review. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e, 39325\u0026ndash;39343. https://doi.org/10.1109/ACCESS.2022.3162594\u003c/li\u003e\n \u003cli\u003eTaherdoost, H. (2024). Insights into Cybercrime Detection and Response: A Review of Time Factor. \u003cem\u003eInformation (Switzerland)\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(5), 273. https://doi.org/10.3390/info15050273\u003c/li\u003e\n \u003cli\u003eTian, T., Gong, Y., Jiang, B., Liu, J., Feng, H., \u0026amp; Lu, Z. (2023). Insider Threat Detection Based On Heterogeneous Graph Neural Network. \u003cem\u003eProceedings - 2023 IEEE 22nd International Conference on Trust, Security and Privacy in Computing and Communications, TrustCom/BigDataSE/CSE/EUC/ISCI 2023\u003c/em\u003e, 628\u0026ndash;635. https://doi.org/10.1109/TrustCom60117.2023.00096\u003c/li\u003e\n \u003cli\u003eTian, T., Zhang, C., Jiang, B., Feng, H., \u0026amp; Lu, Z. (2025). Insider threat detection for specific threat scenarios. \u003cem\u003eCybersecurity\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e(1). https://doi.org/10.1186/s42400-024-00321-w\u003c/li\u003e\n \u003cli\u003eTundis, A., Melnik, M., Naveed, H., \u0026amp; M\u0026uuml;hlh\u0026auml;user, M. (2021). A social media-based over layer on the edge for handling emergency-related events. \u003cem\u003eComputers and Electrical Engineering\u003c/em\u003e, \u003cem\u003e96\u003c/em\u003e, 107570. https://doi.org/10.1016/j.compeleceng.2021.107570\u003c/li\u003e\n \u003cli\u003eWang, Z. Q., \u0026amp; El Saddik, A. (2023a). DTITD: An Intelligent Insider Threat Detection Framework Based on Digital Twin and Self-Attention Based Deep Learning Models. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e11\u003c/em\u003e, 114013\u0026ndash;114030. https://doi.org/10.1109/ACCESS.2023.3324371\u003c/li\u003e\n \u003cli\u003eWang, Z. Q., \u0026amp; El Saddik, A. (2023b). DTITD: An Intelligent Insider Threat Detection Framework Based on Digital Twin and Self-Attention Based Deep Learning Models. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e11\u003c/em\u003e, 114013\u0026ndash;114030. https://doi.org/10.1109/ACCESS.2023.3324371\u003c/li\u003e\n \u003cli\u003eXiao, H., Zhu, Y., Zhang, B., Lu, Z., Du, D., \u0026amp; Liu, Y. (2024). Unveiling shadows: A comprehensive framework for insider threat detection based on statistical and sequential analysis. \u003cem\u003eComputers and Security\u003c/em\u003e, \u003cem\u003e138\u003c/em\u003e, 103665. https://doi.org/10.1016/j.cose.2023.103665\u003c/li\u003e\n \u003cli\u003eYasin, A., Fatima, R., Wen, L., JiangBin, Z., \u0026amp; Niazi, M. (2025). What goes wrong during phishing education? A probe into a game-based assessment with unfavorable results. \u003cem\u003eEntertainment Computing\u003c/em\u003e, \u003cem\u003e52\u003c/em\u003e. https://doi.org/10.1016/j.entcom.2024.100815\u003c/li\u003e\n \u003cli\u003eYuan, S., \u0026amp; Wu, X. (2021a). Deep learning for insider threat detection: Review, challenges and opportunities. \u003cem\u003eComputers \u0026amp; Security\u003c/em\u003e, \u003cem\u003e104\u003c/em\u003e, 102221. https://doi.org/10.1016/j.cose.2021.102221\u003c/li\u003e\n \u003cli\u003eYuan, S., \u0026amp; Wu, X. (2021). Deep learning for insider threat detection: Review, challenges and opportunities. In \u003cem\u003eComputers and Security\u003c/em\u003e (Vol. 104). https://doi.org/10.1016/j.cose.2021.102221\u003c/li\u003e\n \u003cli\u003eYuan, S., \u0026amp; Wu, X. (2021c). Deep learning for insider threat detection: Review, challenges and opportunities. In \u003cem\u003eComputers and Security\u003c/em\u003e (Vol. 104). https://doi.org/10.1016/j.cose.2021.102221\u003c/li\u003e\n \u003cli\u003eZaoui, M., Yousra, B., Yassine, S., Yassine, M., \u0026amp; Karim, O. (2024). A Comprehensive Taxonomy of Social Engineering Attacks and Defense Mechanisms: Toward Effective Mitigation Strategies. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e, 72224\u0026ndash;72241. https://doi.org/10.1109/ACCESS.2024.3403197\u003c/li\u003e\n \u003cli\u003eZhang, C. R., Wang, S., Zhan, D. C., Yu, T. Y., Wang, T. G., \u0026amp; Yin, M. Y. (2021). Detecting Insider Threat from Behavioral Logs Based on Ensemble and Self-Supervised Learning. \u003cem\u003eSECURITY AND COMMUNICATION NETWORKS\u003c/em\u003e, \u003cem\u003e2021\u003c/em\u003e. https://doi.org/10.1155/2021/4148441 WE \u0026nbsp;- Science Citation Index Expanded (SCI-EXPANDED)\u003c/li\u003e\n \u003cli\u003eZhang, C., Wang, S., Zhan, D., Yu, T., Wang, T., \u0026amp; Yin, M. (2021). Detecting Insider Threat from Behavioral Logs Based on Ensemble and Self-Supervised Learning. \u003cem\u003eSecurity and Communication Networks\u003c/em\u003e, \u003cem\u003e2021\u003c/em\u003e, 1\u0026ndash;11. https://doi.org/10.1155/2021/4148441\u003c/li\u003e\n \u003cli\u003eZhu, X., Dong, J., Qi, J., Zhou, Z., Dong, Z., Sun, Y., \u0026amp; Wang, M. (2024). AUTH: An Adversarial Autoencoder Based Unsupervised Insider Threat Detection Scheme for Multisource Logs. \u003cem\u003eIEEE Transactions on Industrial Informatics\u003c/em\u003e, \u003cem\u003e20\u003c/em\u003e(9), 10954\u0026ndash;10965. https://doi.org/10.1109/TII.2024.3393491\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Insider threat detection, social engineering, Deep learning, Behavioral cybersecurity, social manipulation, Human factor, Systematic review, Computer Emergency Response Team Insider Threat Dataset, Explainable artificial intelligence, Cybersecurity strategy","lastPublishedDoi":"10.21203/rs.3.rs-7396895/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7396895/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe exponential expansion of the global digital ecosystem has significantly increased organizational vulnerability to sophisticated insider threat attack vectors. Although Machine Learning and Deep Learning models have improved anomaly detection techniques, a critical gap remains in addressing insider threats influenced by internal social engineering. In particular, Reverse Social Engineering, where malicious insiders manipulate unintentional or innocent colleagues, poses an emerging and underexplored threat. This study systematically reviews forty-nine peer-reviewed articles published between 2015 and April 2025, using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses methodology to investigate current deep learning approaches for insider threat detection. The review highlights a reliance on sequential models such as Long Short-Term Memory and Gated Recurrent Unit algorithms, attention-based transformer models, and graph neural networks. These techniques demonstrate effectiveness in identifying behavioral anomalies and system misuse but fail to detect trust manipulation and social exploitation. Additionally, commonly used datasets, including the Computer Emergency Response Team Insider Threat Dataset from Carnegie Mellon University, DARPA1999, and Enron, do not adequately represent realistic social engineering scenarios, thereby limiting the ability of detection models to address human-driven threats. Traditional evaluation metrics, including Precision, Recall, and F1 Score, also fall short in assessing the contextual and behavioral dimensions of insider threats. This review emphasizes the urgent need for adaptive, context aware and behavior-aware detection frameworks, enriched datasets that incorporate social dynamics, and evaluation models that account for cognitive influence. Addressing these overlooked dimensions is essential for advancing organizational cybersecurity resilience against evolving insider threat landscapes.\u003c/p\u003e","manuscriptTitle":"Deep Learning Models and Social Engineering Dynamics in Insider Threat Detection: A Systematic Review","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-28 09:14:02","doi":"10.21203/rs.3.rs-7396895/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"1249c87e-4eab-433a-a906-bac8f04dd752","owner":[],"postedDate":"August 28th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-09-30T12:23:30+00:00","versionOfRecord":[],"versionCreatedAt":"2025-08-28 09:14:02","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7396895","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7396895","identity":"rs-7396895","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}