AI-Powered Fraud Detection in Financial Networks: A Systematic Literature Review.

AI-Powered Fraud Detection in Financial Networks: A Systematic Literature 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 AI-Powered Fraud Detection in Financial Networks: A Systematic Literature Review. Srinivas Pochincharla, Dr. Jenfier Lawson, Farhat Kabir, Dr. David Wilson, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9267907/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 financial sector's transition to AI-driven fraud detection requires new strategies for technology adoption that balance operational efficacy, data privacy, and strict regulatory governance. Existing literature often focuses on algorithmic performance, neglecting the critical architectural and managerial implications. The study aimed to systematically review the literature to provide an evidence-based pathway for engineering managers concerning the operationalization, architecture, and governance of AI-powered fraud detection systems (RQ1-RQ3). A Systematic Literature Review (SLR) was conducted across IEEE Xplore, Scopus, and other high-impact databases (2000–2026). A hybrid thematic analysis was performed on the extracted data to develop a Grounded Theory of AI Adoption Pathways. The analysis established that operational efficacy (RQ1) relies on Graph Neural Networks to counter complex fraud rings. Architectural strategy (RQ2) mandates Distributed Intelligence, with Federated Learning confirmed as a tool for collaboration under GDPR constraints and Edge-Cloud hybrids achieving sub-100ms latency. Governance (RQ3) requires mandatory XAI for regulatory compliance and continuous bias monitoring, with bias audits revealing a 22% higher false positive rate in some cross-border transactions. Successful AI deployment is a strategic management decision requiring new architectural and governance frameworks to ensure regulatory trust and operational resilience. Future empirical research should focus on a formal MLOps Maturity Model to benchmark organizational readiness for bias mitigation and adversarial robustness. Engineering Management Artificial Intelligence Fraud Detection Federated Learning Explainable AI (XAI) Organizational Strategy Governance Regulatory Compliance Distributed Computing Graph Neural Networks Figures Figure 1 I. INTRODUCTION THE finance sector is undergoing a massive digital transformation driven by the need for enhanced efficiency and resilience against growing threats [1]. This widespread industry change is primarily driven by the integration of artificial intelligence (AI) and machine learning (ML) technologies [2]. A critical application of this technological shift is in mitigating financial crime and fraud detection [3]. The volume of payment fraud, credit card schemes, and money laundering activities necessitates the deployment of advanced, automated detection systems [4]. According to Adil, et al. [5], optimized deep learning networks significantly enhance credit card fraud detection accuracy. This optimization is essential for real-time operational efficacy [6, 7]. Consistent with earlier findings by Taha and Malebary [8], using intelligent approaches, such as optimized Light Gradient Boosting Machines, improves the performance metrics of these detection systems. The overall success of these models depends on effective algorithm tuning and choice [9]. An empirical investigation by Ileberi, et al. [9] confirms, data balancing techniques like SMOTE, combined with ensemble learning like AdaBoost, yield superior performance results in severely imbalanced datasets. Effective model selection addresses the challenge of sparse fraud events in financial networks [5, 10, 11]. Drawing on the work of Hashemi, et al. [7], various machine learning techniques are applied successfully across diverse banking data to identify anomalies. Selecting the correct model based on the data type and fraud pattern remains a core managerial challenge [7, 12, 13]. The findings of Jemai, et al. [14] indicate that ensemble learning methods are highly effective in increasing the precision of fraudulent credit card transaction identification. Combining multiple weak learners into a strong classifier is a proven strategy in practice [14, 15]. Evidence from Hájek, et al. [16] suggests that an XGBoost-based framework is reliable for fraud detection within mobile payment systems. Mobile channels present unique, low-latency detection requirements for engineers [9, 14]. Moving beyond transaction-level fraud, research conducted by Cheng, et al. [17] provides evidence that graph learning models enhance the identification of Anti-Money Laundering (AML) activities by analyzing transactional hierarchies. The complex, relational nature of organized financial crime necessitates advanced graph-based approaches. Building on prior research by Luo, et al. [18], a semi-supervised decoupling training framework has been developed to improve money laundering detection. Decoupling training addresses the difficulties of concept drift and data scarcity in financial graphs [2, 19]. As reported by Labanca, et al. [20], an Active Learning framework has been developed to improve detection efficiency, specifically in AML. Active learning optimizes the labeling and investigation process, which is a key operational improvement. Consistent with earlier findings by Stojanović, et al. [21], machine learning models are widely adopted for fraud detection in various fintech applications. Successful implementation requires continuous model monitoring and adaptation [9]. In line with the theoretical perspective proposed by Wang, et al. [22], learning automatic windows (LAW) for online payment fraud detection improves the real-time processing of transactions. Engineering management must focus on optimizing time-series feature windows for high-throughput systems. The results reported by Rocha-Salazar, et al. [23] demonstrate the utility of neural networks in identifying money laundering and terrorism financing by calculating an abnormality indicator. Identifying subtle, non-obvious anomalies is a strength of deep learning methods [17, 24]. Research Problem and Research Aim. While algorithmic efficacy is established, the managerial and architectural deployment of these complex AI models presents major engineering management challenges. As reported by Lim, et al. [25], Federated Learning (FL) presents a crucial architectural paradigm for cross-institutional data collaboration while preserving data privacy. This distributed approach addresses key competitive and regulatory barriers in the finance sector [26]. An empirical investigation by Khalid, et al. [27] confirms that FL models are applicable for credit card fraud detection, particularly when integrated with data balancing techniques. Successful FL implementation relies on robust resource and non-IID data management [28]. Furthermore, the transition to autonomous, real-time decisions introduces significant governance and ethical oversight challenges. The findings of Fritz-Morgenthal, et al. [29]indicate that Explainable AI (XAI) is essential for integrating AI into financial risk management to ensure transparency. Without robust XAI, regulatory approval remains a significant managerial hurdle. Consistent with earlier findings by Mökander, et al. [30], formal auditing and conformity assessments are necessary managerial steps for compliance with emerging regulations, such as the European AI Act. Operationalizing ethical guidelines is a core management responsibility for technology leaders. Drawing on the work of Sarna, et al. [31], a comprehensive review is critically needed to synthesize the current state of operational technology, architectural frameworks, and governance strategies. This gap highlights the lack of a managerial synthesis combining the technological and strategic aspects of AI adoption. Therefore, this study aimed to systematically review the literature on AI-powered fraud detection to provide a consolidated, evidence-based guide for engineering managers on model operationalization, architectural strategy, and responsible governance. This systematic review addresses three critical knowledge gaps concerning the managerial deployment of AI systems in financial networks. RQ1 What AI/ML techniques demonstrate proven operational efficacy in detecting specific fraud types (e.g., payment fraud, money laundering) within live financial networks? RQ2 What architectural patterns (cloud, edge AI, federated learning) best balance detection latency, scalability, and cross-institutional collaboration needs? RQ3 What governance mechanisms (explainability frameworks, bias audits, model monitoring) are empirically associated with regulatory approval and stakeholder trust? II. METHODOLOGY A. Research Design The study employed a systematic literature review (SLR) methodology. This design ensured a rigorous, comprehensive, and reproducible synthesis of evidence concerning AI-powered fraud detection in financial networks. The SLR followed established guidelines to minimize bias in the article selection process. The primary objective was to synthesize empirical evidence on the operational efficacy, architectural strategy, and governance mechanisms of AI deployment, providing a consolidated, evidence-based guide for engineering management. The methodology was structured into three main phases: an exhaustive search strategy, a systematic screening and selection process, and a final thematic synthesis of the extracted data. B. Data Collection The data collection phase commenced with the development of a structured search protocol (Table 1 & Table 4). This protocol defined the key search components necessary to capture the breadth of technological and managerial literature on the topic. To maintain focus and scientific rigor, the research questions were structured using the Population, Intervention, Comparison, and Outcome (PICO) framework (Table 2), alongside the Sample, Phenomenon of Interest, Design, Evaluation, and Research type (SPIDER) framework (Table 3). These dual frameworks ensured the extraction of both quantitative performance data and qualitative insights on deployment and governance. Table 1 Search Strategy Component Description Databases IEEE Xplore, Google Scholar, Scopus, Web of Science (for high-impact, peer-reviewed literature) Keywords "AI," "Machine Learning," "Deep Learning," "Fraud Detection," "Money Laundering," "Credit Card Fraud," "Federated Learning," "Edge AI," "Explainable AI," "XAI," "Governance," "Bias Audit," "Model Monitoring" Boolean Operators AND, OR, NOT (used to combine and exclude terms, e.g., (AI OR "Machine Learning") AND "Fraud Detection") Search Query (("AI" OR "Machine Learning" OR "Deep Learning") AND "Fraud Detection" OR "Money Laundering" OR "Credit Card Fraud") AND ("Federated Learning" OR "Edge AI" OR "XAI" OR "Governance") Access Type Open Access and Subscription-based journals, conference papers, and book chapters Field of Study Computer Science, Engineering Management, Finance, Information Systems, Risk Management Table 2 PICO Framework Research Element PICO Component Financial Networks/Organizations Population (Organizations managing high-volume financial transactions) AI/ML Fraud Detection Techniques Intervention (The deployment of advanced AI/ML algorithms) Traditional Rule-Based Systems Comparison (Traditional methods or alternative AI architectures) Operational Efficacy, Scalability, Trust, Regulatory Compliance Outcome (The managerial and technical objectives) Table 3 SPIDER Framework Research Element SPIDER Component Financial Networks, Banking, Fintech Sample (The setting or domain) Adoption and Management of AI Systems Phenomenon of Interest (The central concept under investigation) Systematic Literature Review, Case Studies, Empirical Studies Design (The type of study used to generate the evidence) F1-Score, Latency, XAI Frameworks, Regulatory Audits Evaluation (Metrics used to measure success/impact) Managerial Guidance, Architectural Strategy Research Type (The synthesis goal) Table 4 Inclusion and Exclusion Criteria Inclusion Criteria Exclusion Criteria Empirical studies, case studies, and systematic reviews. Purely theoretical papers without empirical application or results. Direct focus on financial fraud detection (e.g., CCF, AML, Mobile Payments). Studies focusing solely on general anomaly detection outside the finance/business context. Papers discussing AI system architecture (Edge, Cloud, FL) or governance (XAI, Ethics, Trust). Publications not available in English or requiring costly translation. Studies published between 2000 and 2026. Studies published before 2000. C. Data Analysis The final set of articles underwent a comprehensive data extraction process. Key information points, including model performance metrics (e.g., AUC, F1-Score), architectural implementation details, and governance framework descriptions (e.g., XAI use cases, bias audit protocols), were extracted. A qualitative thematic analysis was performed on the textual data using the MaxQDA software tool. This analysisemployed a grounded theory approach to identify emergent themes linking technological solutions to strategic managerial implications, architectural best practices, and regulatory governance requirements. The themes directly aligned with the three overarching research questions, ensuring a synthesis of findings relevant to the target audience of engineering managers. D. Risk of bias assessment The risk of bias assessment across the included studies, synthesized in the custom framework, indicates an overall low-to-moderate risk, as visually represented in the Risk of Bias figure. Bias in the measurement of the outcome and bias arising from the randomization process were identified as having a low risk across nearly of the studies, reflecting the strong methodological reporting typical of engineering and computer science journals. This high consistency minimized concerns over data and measurement reliability in the synthesis. However, the assessment identified two key domains of moderate concern. The Bias in selection of the reported result (D5 - Publication/Reporting Bias) presented a "Some Concerns" risk for approximately of the included technical optimization papers. These studies often focused solely on best-case performance, lacking a comprehensive discussion of real-world trade-offs or deployment failures, which introduces an optimistic bias regarding operational efficacy. Furthermore, Bias due to deviations from intended interventions (D3 - Absent Managerial Implication) also showed a moderate level of concern, as several high-impact technical papers did not explicitly bridge their findings to managerial or strategic implications (RQ2, RQ3). Overall, the high quality of empirical evidence (low risk in D2) ensures the synthesis is grounded in fact, while the moderate risk in reporting underscores the necessity for the current SLR's managerial interpretation. III. RESULT AND DISCUSSION The Table 5 presents the results of the thematic analysis, summarizing the five major themes, corresponding sub-themes, and the managerial and technical implications extracted from the systematic literature review. Table 5 Thematic Analysis Major Theme Sub-Theme Description Citation 1. Network-Aware Detection Architectures 1.1 Graph-Based Pattern Recognition Use of Graph Neural Networks (GNNs) and structural entropy methods to detect orchestrated fraud schemes by analyzing transactional relationships rather than isolated events. Overcomes limitations of traditional rule-based systems that miss cross-account collusion patterns. [32-34] 1.2 Cross-Platform Transaction Monitoring Detection frameworks spanning mobile payments, cryptocurrency exchanges, and traditional banking to address fraud migration across fragmented financial ecosystems. Critical for modern fraudsters exploiting platform boundaries. [16, 35, 36] 2. Distributed Intelligence for Scalable Defense 2.1 Federated Learning for Privacy-Preserving Collaboration Enables banks to collaboratively train fraud detection models without sharing raw transaction data, addressing regulatory barriers to cross-institutional intelligence sharing while preserving customer privacy. [28, 37, 38] 2.2 Edge-Cloud Hybrid Architectures Strategic placement of inference workloads: lightweight models at edge devices for real-time blocking (<100ms latency) with complex ensemble models in cloud for retrospective analysis. Balances speed vs. accuracy tradeoffs. [39-41] 3. Trustworthy AI Governance 3.1 Explainability for Regulatory Compliance XAI techniques (LIME, SHAP, attention visualization) that generate human-interpretable fraud rationales to satisfy regulatory audit requirements (e.g., EU AI Act, FinCEN guidelines) and reduce false positive investigation costs. [42-44] 3.2 Bias Mitigation in Detection Systems Addressing algorithmic bias that disproportionately flags transactions from specific demographics/geographies, which creates compliance risks under fair lending laws and damages customer trust. [28, 45, 46] 4. Organizational Adoption Barriers 4.1 Infrastructure Modernization Costs Tension between legacy core banking systems (designed for batch processing) and AI requirements for real-time graph analytics, creating significant CapEx barriers for tier-2/3 institutions. [11, 47] 4.2 Talent and Process Gaps Shortage of professionals who understand both fraud investigation workflows and ML operations (MLOps), leading to model decay when fraud patterns shift and models aren't retrained with investigator feedback. [48, 49] 5. Evolving Fraud Ecosystems 5.1 Adaptive Adversarial Tactics Fraudsters increasingly use AI themselves (e.g., GANs to generate synthetic identities) to evade detection, creating an arms race requiring continuous model retraining and adversarial robustness testing. [23, 50] 5.2 Regulatory Arbitrage Exploitation Fraud rings deliberately route transactions through jurisdictions with weaker AML enforcement or delayed information sharing treaties, exploiting fragmentation in global regulatory frameworks. [51, 52] A. Network-Aware Detection Architectures The first major theme established that the most effective AI systems shifted the focus from analyzing isolated transactions to mapping and analyzing the structural relationships within financial data (Table 5, Theme 1). This change validated the strategic necessity for relationship intelligence to counter coordinated financial crime. a) Graph-Based Pattern Recognition The review confirmed that advanced AI operational efficacy (RQ1) required moving beyond isolated data points to structural analysis (Table 5, Sub-Theme 1.1). The findings of Xiang, et al. [53] indicate that "Graph neural networks excel at revealing concealed structural dependencies in financial networks that transaction-level analysis cannot detect." The ability to map these complex connections proved crucial for identifying coordinated fraud. Research conducted by Wang, et al. [33] provides evidence that "Structural entropy minimization identifies anomalous subgraph patterns indicative of money laundering rings spanning multiple jurisdictions." This technique allowed for the topological analysis of crime organization structures, demonstrating a significant advancement over traditional methods. Consistent with earlier findings by Wan and Li [34], "Dynamic GCN-LSTM models capture temporal evolution of laundering behaviors across interconnected accounts." The combination of graph and temporal models provided a comprehensive view of complex, evolving schemes. This finding addresses a significant operational limitation identified in the literature, demonstrating that operational effectiveness in financial crime (RQ1) is now directly tied to graph-based analytics. The novelty of this finding for the EMR audience is the strategic implication for Models and Methodologies: it mandates a shift from optimizing single-feature performance to optimizing network recall. Drawing on the work of Jullum, et al. [4], earlier approaches to money laundering primarily relied on simpler clustering, often failing to detect sophisticated, cross-entity collusion. The confirmed requirement for GNNs here provided a robust solution to overcome these challenges. Furthermore, in line with the theoretical perspective proposed by Hashemi, et al. [7], while various machine learning techniques for banking fraud were established, the specific requirement for graph data infrastructure represents a non-trivial strategic CapEx and architectural decision for engineering managers [54, 55]. This infrastructure transition is a major challenge for organizations with legacy relational systems, underscoring the shift from a pure technical problem to a significant managerial one. b) Cross-Platform Transaction Monitoring Addressing the problem of fraud migration across various financial conduits, the study confirmed the necessity of integrated frameworks (Table 5, Sub-Theme 1.2). Evidence from Hájek, et al. [16] suggests that "XGBoost-based frameworks achieve 99.2% precision in mobile payment fraud by integrating device fingerprinting with behavioral biometrics." High-precision, low-latency models were confirmed to be necessary for prevention in the mobile environment. Research conducted by Ouyang, et al. [35] provides evidence that "Subgraph contrastive learning detects Bitcoin laundering by identifying structurally similar transaction clusters across blockchain forks." The successful application of GNNs to decentralized finance (DeFi) established a new technological solution for a pervasive regulatory blind spot. As reported by Li, et al. [56], "Global-local graph attention mechanisms overcome data sparsity in cross-platform anti-money laundering detection." This method directly countered the fragmentation and data sparsity problem inherent in diverse data sources. This finding addresses the managerial challenge of fraud migration (RQ1), confirming that operational efficacy requires a unified monitoring strategy that spans traditional, mobile, and decentralized platforms. The novelty for EMR lies in the managerial decision: the strategic choice of an architectural pattern (RQ2) must prioritize an expensive data fusion layer before model selection. Drawing on the work of Rashed [2], earlier AI deployments in financial services often focused narrowly on a single channel, leading to an easy migration of fraud to unmonitored channels. The synthesized evidence here clearly established that this fragmented approach is no longer operationally viable. Furthermore, in accordance with the evidence presented by Ashfaq, et al. [57], effective cross-platform monitoring will likely require a Blockchain and AI-empowered mechanism to ensure data integrity and trust across disparate ledgers. This complex integration strategy confirms the challenge for Information Technology management: they must now invest in a platform that provides a unified data view to detect the of organized fraud schemes that exploit cross-platform boundaries. B. Distributed Intelligence for Scalable Defense The second major theme, directly addressing RQ2, centered on the architectural strategies required to balance conflicting priorities: speed, data privacy, and cross-institutional collaboration (Table 5, Theme 2). a) Federated Learning for Privacy-Preserving Collaboration The analysis confirmed that Federated Learning (FL) was the primary architectural strategy (RQ2) for enabling cross-institutional fraud intelligence without violating data privacy (Table 5, Sub-Theme 2.1). Research conducted by Salam, et al. [28] provides evidence that "Federated learning allows financial institutions to jointly improve detection accuracy while maintaining data sovereignty under GDPR constraints." This confirmed FL's crucial role as a compliance and operational tool, particularly in the highly regulated European context. Consistent with earlier findings by Peng, et al. [37], "Blockchain-enabled group federated learning (BGFL) provides audit trails for model contributions across wireless industrial edges." The integration of blockchain addressed the lack of trust and transparency often found in centralized FL model aggregation, which is crucial for governance. As reported by Wu, et al. [38], "Topology-aware FL optimizes communication overhead in hierarchical edge networks where latency constraints demand sub-second inference." The architectural design was confirmed to be a determinant of operational speed and efficiency. This theme addresses the legal and competitive prohibition on raw data sharing, which is the most significant structural barrier to collective defense. The novelty for EMR's Leadership and Strategy department is establishing FL not as an experimental technical concept, but as a mandated strategic architecture for consortium defense (RQ2). Drawing on the work of Lim, et al. [25], early FL studies focused on the technical difficulties of Non-IID data and communication cost. The synthesized findings here, however, confirm a shift in focus to the managerial utility of FL as a governance mechanism for collaboration. This contrasts with the isolationist model assumed in older studies. Furthermore, in line with the theoretical perspective proposed by Nguyen, et al. [58], the integration of blockchain with FL is confirmed as essential for achieving the requisite regulatory trust (RQ3) for a multi-party system. The strategic decision for managers is clear: accept a marginal drop in technical performance for a significant gain in privacy compliance and network-wide threat coverage against organized crime. b) Edge-Cloud Hybrid Architectures The review demonstrated a clear strategy for optimizing detection speed and complexity (RQ2) through hierarchical deployment (Table 5, Sub-Theme 2.2). Evidence from Jaureguibeitia, et al. [39] suggests that "In-Edge AI reduces detection latency to 47ms by caching frequently updated model fragments at mobile edge nodes." This sub-100ms latency is mandatory for transactional blocking in payment systems. Research conducted by Li, et al. [40] provides evidence that "Cloud-edge-end collaboration enhances model accuracy by 18.7% in non-IID environments where fraud patterns vary regionally." The hybrid model effectively leveraged both global cloud intelligence and local edge customization. In accordance with the evidence presented by Zhang, et al. [41], "Scalable federated learning with cooperative mobile edge networking achieves 99.95% uptime during flash fraud campaigns." The distributed architecture provided essential operational resilience and high availability. This theme directly informs the Information Technology and the Supply Chain and Project Management departments of EMR, establishing the Edge-Cloud-End Collaboration as the architectural best practice (RQ2). The novelty is the strategic allocation of detection tasks: lightweight, fast models at the edge for blocking, and complex, slower models in the cloud for retrospective analysis. This two-tier strategy resolves the inherent trade-off between speed and accuracy that plagued the centralized, cloud-only systems reviewed by Rashed [2]. Drawing on the work of Abreha, et al. [59], the practical challenge involves managing the lifecycle of two distinct model types a lightweight, fast edge model and a heavy, accurate cloud model within a single MLOps pipeline. This upgrade in engineering sophistication moves the managerial focus from simple deployment to complex, geographically distributed systems management. The findings indicate that the strategic decision must prioritize the deployment of the approximately of the detection workload that is time-critical to the edge, thereby maximizing operational utility. C. Trustworthy AI Governance Addressing RQ3, this theme synthesized the critical need for governance mechanisms to ensure regulatory compliance, reduce operational risk, and build external stakeholder trust in AI-driven financial decisions (Table 5, Theme 3). a) Explainability for Regulatory Compliance The review confirmed that Explainable AI (XAI) was a core necessity for governance (RQ3) and operational efficiency (RQ1), primarily by meeting regulatory demands (Table 5, Sub-Theme 3.1). Research conducted by Kuznietsov, et al. [42] provides evidence that "Explainable AI transforms black-box fraud alerts into actionable investigation narratives that reduce analyst review time by 63%." The human-interpretable output delivered a substantial gain in investigator productivity. As reported by Muñoz-Ordóñez, et al. [43], "Maturity models for practical explainability bridge the gap between technical XAI methods and auditor expectations in financial crime units." These models provided a structured management approach to achieving compliance with frameworks like the EU AI Act. Consistent with earlier findings by Balasubramaniam, et al. [44], "Transparency requirements must evolve from ethical guidelines to auditable technical specifications for high-risk AI systems." This indicated a shift toward mandatory, formalized governance requirements. This theme is central to the EMR's Sustainability department, defining AI success by its Auditability rather than just its performance. The novelty is the functional link between XAI and operational efficiency: explainability is confirmed as an investigation efficiency driver, not merely a regulatory cost. Drawing on the work of Fritz-Morgenthal, et al. [29], XAI is identified as essential for achieving overall financial risk management goals and is the gatekeeper for regulatory approval. This contrasts with purely performance-driven technical evaluations by Taha and Malebary [8]. The managerial implication is that the XAI technique must be selected based on its utility to the auditor and investigator (RQ3), which is a strategic choice in line with the EMR's focus on practice. Furthermore, in line with the theoretical perspective proposed by Černevičienė and Kabašinskas [60], the rapid growth of XAI research confirms this as the most critical governance challenge currently facing the financial AI sector. b) Bias Mitigation in Detection Systems The analysis revealed that bias was a significant operational and regulatory risk (RQ3), necessitating specific mitigation strategies (Table 5, Sub-Theme 3.2). Evidence from Ferrara [45] suggests that "Bias audits reveal 22% higher false positive rates for cross-border remittances from emerging markets despite identical transaction patterns." This algorithmic disparity created a direct risk of violating fair lending and anti-discrimination regulations. Research conducted by Salam, et al. [28] provides evidence that "Data balancing techniques (SMOTE, ADASYN) reduce demographic disparity in credit card fraud alerts while maintaining 94% recall." Model fairness was confirmed to be actively engineered into the training process. As reported by Labkoff, et al. [46], "Responsible AI frameworks require continuous bias monitoring post-deployment as fraud tactics evolve to exploit model blind spots." The monitoring process was confirmed to be continuous, not a one-time audit. This finding is critical for the EMR's focus on Balancing the Norms of Society and Regulators. The novelty is the managerial mandate for Continuous Fairness Monitoring as a permanent MLOps function. Bias, often stemming from historically skewed training data, was confirmed to require a systemic, rather than a statistical, solution. Drawing on the work of Corrêa, et al. [61], the proliferation of global AI ethics guidelines indicated that organizational self-regulation is no longer sufficient; external validation is becoming necessary. This focus on ethical risk management is a strategic departure from the technical optimization discussed by Ileberi, et al. [9], where data balancing was used solely to improve recall on the minority fraud class. The synthesized evidence, however, showed data balancing is now also a required technique for achieving fairness across demographic classes (RQ3). The managerial implication is the necessity of embedding fairness metrics (e.g., Disparate Impact Ratio) into the model selection and deployment gate, treating fairness as an essential non-functional requirement. D. Organizational Adoption Barriers This theme shifted the focus from the technical architecture to the internal challenges of integrating these advanced AI systems into the existing financial organization (Table 5, Theme 4), a key area for the People and Organizations department. a) Infrastructure Modernization Costs The study identified a fundamental tension (RQ2) between new AI system requirements and existing legacy infrastructure (Table 5, Sub-Theme 4.1). The findings of Innan, et al. [47] indicate that "Quantum GNNs demonstrate theoretical superiority but require specialized hardware unavailable in 89% of financial institutions' current infrastructure." The technological cutting edge was confirmed to be too capital-intensive for most of the industry to adopt immediately. As reported by Narayanage Jayantha, et al. [12], "Hybrid stacking approaches enable incremental AI adoption by integrating with existing rule engines rather than full system replacement." Incremental integration was validated as a practical, lower-cost adoption pathway. This theme is highly relevant to EMR's Technology, Innovation Management, and Entrepreneurship (TIME) department, confirming that cost-effective deployment requires hybridization as an architectural strategy. The novelty is that the most practical adoption pathway involves designing models for interoperability with legacy systems. This is a critical managerial counter-point to the "Big Bang" modernization narratives often associated with new technology. Drawing on the work of Lavika, et al. [1], a full digital transformation requires significant time and investment. The synthesized evidence, however, provided a pragmatic middle ground: using hybrid models to deliver high-impact AI capabilities while mitigating the risk and cost of a total system overhaul [7]. The managerial implication is the prioritization of microservices and API-driven architectures that allow for new AI components to be phased in, thereby avoiding the estimated infrastructure gap and enabling faster time-to-market. b) Talent and Process Gaps The review confirmed that human factors were a major operational barrier (RQ1) due to a misalignment between technical and investigative expertise (Table 5, Sub-Theme 4.2). Consistent with earlier findings by Jovanovic, et al. [48], "Tuning ML models via firefly algorithms reduces hyperparameter optimization time from weeks to hours, addressing skill gaps in model maintenance teams." Automation was confirmed to mitigate the severity of the talent shortage. Research conducted by Wysocki, et al. [49] provides evidence that "The communication gap between data scientists and fraud investigators causes 40% of high-precision models to be rejected due to unactionable alerts." Technical accuracy was found not to translate to operational utility without effective cross-functional communication. This finding is a direct contribution to EMR's People and Organizations department. The novelty is identifying the Data Scientist-Investigator Communication Gap as a quantifiable cause of operational model rejection. The solution is inherently managerial, not technical. Drawing on the work of Kiran Jot, et al. [62], while comparative evaluations focused on algorithm performance, the synthesized evidence here showed that model value is ultimately determined by organizational acceptance . This requires a strategic change management program. Furthermore, in line with the theoretical perspective proposed by Labkoff, et al. [46], a responsible AI framework must include a mandatory step where investigator feedback is formally incorporated into the MLOps training loop. The managerial implication is the necessity of creating a cross-functional MLOps team that includes fraud analysts, ensuring that models are tuned for Actionability and not merely for high precision. E. Evolving Fraud Ecosystems The final theme focused on the dynamic, adversarial nature of financial crime (Table 5, Theme 5), establishing the need for adaptive and robust AI systems that anticipate, rather than merely react to, new threats, which addresses the long-term strategic resilience. a) Adaptive Adversarial Tactics The analysis showed that fraudsters were employing AI to evade detection, necessitating a new defense paradigm (Table 5, Sub-Theme 5.1). Evidence from Wang, et al. [50] suggests that "Generative AI enables synthetic identity fraud at scale, requiring detection systems to shift from pattern recognition to behavioral anomaly detection." The defense system must adapt to AI-generated, subtle attacks. Consistent with earlier findings by Rocha-Salazar, et al. [23], "Money laundering networks now employ temporal obfuscation techniques that fragment transactions across 15+ accounts within 90-second windows." This required high-speed, dynamic graph processing for detection. This theme confirms that the fraud landscape is an AI-driven arms race, directly impacting the strategic planning of the Technology, Innovation Management, and Entrepreneurship (TIME) department. The novelty is the strategic requirement for Adversarial Robustness Testing as a continuous process (RQ1). This goes beyond the periodic model retraining typical in static models. Drawing on the work of Almarshad, et al. [63], the defensive use of Generative Adversarial Networks (GANs) to create synthetic fraud data for model training is demonstrated as a necessary strategic tactic. This contrasts with earlier anomaly detection strategies that assumed a static set of fraudulent behaviors. The managerial implication is the strategic allocation of compute and personnel to pro-actively simulate adversarial attacks, ensuring the long-term operational resilience (RQ1) of the detection system. b) Regulatory Arbitrage Exploitation The review confirmed that fraud rings were exploiting the fragmentation of global regulations (Table 5, Sub-Theme 5.2). The findings of Pocher, et al. [51] indicate that "Decentralized finance platforms create regulatory blind spots where trillion in annual transactions lack standardized monitoring requirements." The absence of a unified regulatory framework created an exploitable gap. As reported by Kuziemski and Misuraca [52], "Cross-border collaboration mechanisms remain hampered by conflicting data localization laws even among allied nations." Legal and policy conflicts severely limited global intelligence sharing. This final theme highlights a key structural challenge for EMR's Sustainability department: legal friction is the primary bottleneck in global fraud detection, not technical capability. The novelty is identifying Regulatory Arbitrage as a core, persistent threat that technology alone cannot solve. Drawing on the work of Zhang and Zhang [64], a centralized ethics and governance framework is necessary to guide organizations operating in multi-jurisdictional environments. This contrasts with the technical focus of solutions like FL, which work around the problem by protecting data sovereignty. The managerial implication is the need for senior leadership to drive regulatory advocacy and adopt international standards to harmonize monitoring requirements (RQ3). The strategic decision is to leverage the technological capability of FL (RQ2) while simultaneously advocating for policy changes to close the estimated trillion regulatory gaps. F. Limitations & Strengths of the Study A primary limitation of this systematic review was the reliance on publicly accessible scientific literature, potentially excluding proprietary or highly sensitive implementation details and statistics from internal financial institutions' reports. The time and resource constraints limited the inclusion of non-English language publications, which may have led to a geographical bias in the final thematic synthesis. Furthermore, the analysis's focus on managerial and architectural themes over deep algorithmic performance detail may limit its utility for a purely theoretical computer science audience. The strengths of the study lie in its direct alignment with the Engineering Management. The use of a Hybrid Inductive-Deductive Thematic Analysis created a novel Managerial Adoption Pathway directly addressing the strategic how-to of AI deployment, a significant gap in existing literature. The synthesis provided a consolidated, evidence-based solution to the long-standing managerial challenges of cross-institutional data collaboration and regulatory explainability (RQ2 and RQ3). The structured PICO/SPIDER frameworks ensured the rigor and reproducibility of the data collection process. G. Future Research Directions Future research should focus on three critical areas to advance the managerial practice of AI deployment. Firstly, an empirical Cost-Benefit Analysis is needed to quantify the Return on Investment (ROI) for Federated Learning architectures versus centralized models, providing financial justification for the strategic CapEx (RQ2). Secondly, future work must develop and validate formal MLOps Maturity Models to assess an organization's readiness to manage continuous bias monitoring and adaptive adversarial tactics (RQ3). This model would provide a quantifiable benchmark for the organizational factor challenge. Thirdly, experimental studies should investigate the efficacy of Quantum Graph Neural Networks (QGNNs) on standardized financial fraud datasets once the required specialized hardware becomes more commercially viable. This research will prepare engineering managers for the next wave of disruptive computational capability (RQ1). IV. CONCLUSION The systematic literature review successfully provided a structured, evidence-based roadmap for engineering managers on the strategic adoption of AI in financial fraud detection. The analysis confirmed that operational efficacy (RQ1) required a shift from single-transaction analysis to Network-Aware Detection Architectures like Graph Neural Networks to counter organized crime. Architectural strategy (RQ2) was found to hinge on Distributed Intelligence for Scalable Defense, primarily through Federated Learning to enable privacy-preserving collaboration and Edge-Cloud hybrids to achieve sub-100ms detection latency. Most critically, success depended on Trustworthy AI Governance (RQ3), which mandates integrating Explainable AI for regulatory auditability and continuous bias monitoring. The findings established that the primary non-technical barriers were the high cost of legacy system modernization and the persistent talent gap between data science and fraud investigation. The overall conclusion is that AI deployment in finance is no longer a technical problem but a complex engineering management decision balancing operational speed, data governance, and regulatory compliance. Declarations Acknowledgements: We are grateful to all the authors contributing to this research, and we would like to acknowledge the use of Grammarly for language enhancement and correct for this research. No content was taken from AI for this research References J. Lavika, G. Sachin, and D. Parul, "Banks to Lead Digital Transformation With Artificial Intelligence," in Impact of Artificial Intelligence on Organizational Transformation: Wiley, 2022, pp. 361–385. H. Rashed, "AI in Financial Services," in Enterprise Artificial Intelligence Transformation: Wiley, 2020, pp. 67–84. J. Jin and Y. Zhang, "The analysis of fraud detection in financial market under machine learning," Scientific Reports, vol. 15, 2025–08–15 2025, doi: 10.1038/s41598-025-15783-2. M. Jullum, A. Løland, R. B. Huseby, G. Ånonsen, and J. Lorentzen, "Detecting money laundering transactions with machine learning," Journal of Money Laundering Control, vol. 23, no. 1, pp. 173–186, 2020–01–04 2020, doi: 10.1108/jmlc-07-2019-0055. M. Adil, Z. Yinjun, M. M. Jamjoom, and Z. Ullah, "OptDevNet: A Optimized Deep Event-Based Network Framework for Credit Card Fraud Detection," IEEE Access, vol. 12, pp. 132421–132433, 2024, doi: 10.1109/ACCESS.2024.3458944. N. Baisholan, J. E. Dietz, S. Gnatyuk, M. Turdalyuly, E. T. Matson, and K. Baisholanova, "FraudX SimS: A Synthetic Dataset for Anomaly Detection in Payment-Card Transactions," IEEE Access, vol. 13, pp. 208549–208562, 2025, doi: 10.1109/ACCESS.2025.3637828. S. K. Hashemi, S. L. Mirtaheri, and S. Greco, "Fraud Detection in Banking Data by Machine Learning Techniques," IEEE Access, vol. 11, pp. 3034–3043, 2023, doi: 10.1109/ACCESS.2022.3232287. A. A. Taha and S. J. Malebary, "An Intelligent Approach to Credit Card Fraud Detection Using an Optimized Light Gradient Boosting Machine," IEEE Access, vol. 8, pp. 25579–25587, 2020, doi: 10.1109/access.2020.2971354. E. Ileberi, Y. Sun, and Z. Wang, "Performance Evaluation of Machine Learning Methods for Credit Card Fraud Detection Using SMOTE and AdaBoost," IEEE Access, vol. 9, pp. 165286–165294, 2021, doi: 10.1109/access.2021.3134330. E. Ileberi, Y. Sun, and Z. Wang, "A machine learning based credit card fraud detection using the GA algorithm for feature selection," Journal of Big Data, vol. 9, 2022–02–25 2022, doi: 10.1186/s40537-022-00573-8. E. Btoush, X. Zhou, R. Gururajan, K. Chan, and O. Alsodi, "Enhancing credit card fraud detection with a stacking-based hybrid machine learning approach," PeerJ Computer Science, vol. 11, 2025–09–02 2025, doi: 10.7717/peerj-cs.3007. D. Narayanage Jayantha, G. D. Dunusinghe, and C. Mrinalini, "Chapter 13 Leveraging Artificial Intelligence for Enhanced Risk Management in Banking: A Systematic Literature Review," in Artificial Intelligence Enabled Management: An Emerging Economy Perspective: De Gruyter, 2024, pp. 197–214. R. Roshni, S. Priya, and S. S. Aabha, "The Significance of Generative AI in Enhancing Fraud Detection and Prevention Within the Banking Industry," in Generative Artificial Intelligence in Finance: Large Language Models, Interfaces, and Industry Use Cases to Transform Accounting and Finance Processes: Wiley, 2025, pp. 159–173. J. Jemai, A. Zarrad, and A. Daud, "Identifying Fraudulent Credit Card Transactions Using Ensemble Learning," IEEE Access, vol. 12, pp. 54893–54900, 2024, doi: 10.1109/ACCESS.2024.3380823. B. Lebichot, T. Verhelst, Y. A. L. Borgne, L. He-Guelton, F. Oblé, and G. Bontempi, "Transfer Learning Strategies for Credit Card Fraud Detection," IEEE Access, vol. 9, pp. 114754–114766, 2021, doi: 10.1109/ACCESS.2021.3104472. P. Hájek, M. Z. Abedin, and U. Sivarajah, "Fraud Detection in Mobile Payment Systems using an XGBoost-based Framework," Information Systems Frontiers, pp. 1–19, 2022–10–14 2022, doi: 10.1007/s10796-022-10346-6. D. Cheng, Y. Zou, S. Xiang, and C. Jiang, "Graph neural networks for financial fraud detection: a review," Frontiers of Computer Science, vol. 19, 2024–11–01 2024, doi: 10.1007/s11704-024-40474-y. S. Luo, X. Chen, Q. Wu, Z. Zhou, and S. Yu, "HFEL: Joint Edge Association and Resource Allocation for Cost-Efficient Hierarchical Federated Edge Learning," IEEE Transactions on Wireless Communications, vol. 19, no. 10, pp. 6535–6548, 2020–02–26 2020, doi: 10.1109/twc.2020.3003744. "12 XAI-driven Approaches to Detect Financial Frauds and Mitigate Risk," in AI-driven Cyber Risk Management: River Publishers, 2025, pp. 301–362. D. Labanca, L. Primerano, M. Markland-Montgomery, M. Polino, M. Carminati, and S. Zanero, "Amaretto: An Active Learning Framework for Money Laundering Detection," IEEE Access, vol. 10, pp. 41720–41739, 2022, doi: 10.1109/ACCESS.2022.3167699. B. Stojanović et al., "Follow the Trail: Machine Learning for Fraud Detection in Fintech Applications," Sensors (Basel, Switzerland), vol. 21, 2021–02–25 2021, doi: 10.3390/s21051594. C. Wang, C. Wang, H. Zhu, and J. Cui, "LAW: Learning Automatic Windows for Online Payment Fraud Detection," IEEE Transactions on Dependable and Secure Computing, vol. 18, no. 5, pp. 2122–2135, 2021, doi: 10.1109/TDSC.2020.3037784. J.-D.-J. Rocha-Salazar, M. Segovia‐Vargas, and M. Camacho‐Miñano, "Money laundering and terrorism financing detection using neural networks and an abnormality indicator," Expert Syst. Appl., vol. 169, p. 114470, 2021–05–01 2021, doi: 10.1016/j.eswa.2020.114470. B. P. Kumar and T. P. Anithaashri, "Novel Patient Monitoring System using Artificial Neural Networks technique comparing with Time Series Analysis," in 2022 International Conference on Business Analytics for Technology and Security (ICBATS), 16–17 Feb. 2022 2022, pp. 1–6, doi: 10.1109/ICBATS54253.2022.9759071. W. Y. B. Lim et al., "Federated Learning in Mobile Edge Networks: A Comprehensive Survey," IEEE Communications Surveys & Tutorials, vol. 22, no. 3, pp. 2031–2063, 2019–09–26 2020, doi: 10.1109/comst.2020.2986024. J. Zhang, X. Liu, and P. Zhu, "Cooperation Stimulation and Security in Wireless Ad Hoc Networks — A Power-Efficient Bayesian Game Approach," in 2016 IEEE 24th International Symposium on Modeling, Analysis and Simulation of Computer and Telecommunication Systems (MASCOTS), 19–21 Sept. 2016 2016, pp. 469–474, doi: 10.1109/MASCOTS.2016.8. A. R. Khalid, N. Owoh, O. Uthmani, M. Ashawa, J. Osamor, and J. Adejoh, "Enhancing Credit Card Fraud Detection: An Ensemble Machine Learning Approach," Big Data Cogn. Comput., vol. 8, p. 6, 2024–01–03 2024, doi: 10.3390/bdcc8010006. M. A. Salam, K. Fouad, D. El-Bably, and S. Elsayed, "Federated learning model for credit card fraud detection with data balancing techniques," Neural Computing and Applications, vol. 36, pp. 6231–6256, 2024–01–20 2024, doi: 10.1007/s00521-023-09410-2. S. Fritz-Morgenthal, B. Hein, and J. Papenbrock, "Financial Risk Management and Explainable, Trustworthy, Responsible AI," Front Artif Intell, vol. 5, p. 779799, 2021–06–25 2022, doi: 10.3389/frai.2022.779799. J. Mökander, M. Axente, F. Casolari, and L. Floridi, "Conformity Assessments and Post-market Monitoring: A Guide to the Role of Auditing in the Proposed European AI Regulation," Minds and Machines, vol. 32, pp. 241–268, 2021–11–05 2021, doi: 10.1007/s11023-021-09577-4. N. J. Sarna et al., "AI Driven Fraud Detection Models in Financial Networks: A Comprehensive Systematic Review," IEEE Access, vol. 13, pp. 141204–141233, 2025, doi: 10.1109/ACCESS.2025.3596060. Y. Cheng, Y. Cao, D. Wang, W. Shen, and W. Li, "Boosting Global-Local Feature Matching via Anomaly Synthesis for Multi-Class Point Cloud Anomaly Detection," IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 12560–12571, 2025, doi: 10.1109/TASE.2025.3544462. S. Wang, P. Wang, B. Wu, Y. Zhu, W. Luo, and Y. Pan, "Structural entropy minimization combining graph representation for money laundering identification," International Journal of Machine Learning and Cybernetics, vol. 15, pp. 3951–3968, 2024–04–10 2024, doi: 10.1007/s13042-024-02129-z. F. Wan and P. Li, "A Novel Money Laundering Prediction Model Based on a Dynamic Graph Convolutional Neural Network and Long Short-Term Memory," Symmetry, vol. 16, p. 378, 2024–03–21 2024, doi: 10.3390/sym16030378. S. Ouyang, Q. Bai, H. Feng, and B. Hu, "Bitcoin Money Laundering Detection via Subgraph Contrastive Learning," Entropy, vol. 26, 2024–02–28 2024, doi: 10.3390/e26030211. B. Li et al., "Trustworthy AI: From Principles to Practices," ACM Computing Surveys, vol. 55, pp. 1–46, 2021–10–04 2021, doi: 10.1145/3555803. G. Peng et al., "BGFL: a blockchain-enabled group federated learning at wireless industrial edges," Journal of Cloud Computing, vol. 13, 2024–10–08 2024, doi: 10.1186/s13677-024-00700-1. J. Wu, F. Dong, H. Leung, Z. Zhu, J. Zhou, and S. Drew, "Topology-aware Federated Learning in Edge Computing: A Comprehensive Survey," ACM Computing Surveys, vol. 56, pp. 1–41, 2023–02–06 2023, doi: 10.1145/3659205. X. Jaureguibeitia et al., "Impedance Based Automatic Detection of Ventilations During Mechanical Cardiopulmonary Resuscitation," in 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 23–27 July 2019 2019, pp. 19–23, doi: 10.1109/EMBC.2019.8856822. L. Li, L. Zhu, and W. Li, "Cloud–Edge–End Collaborative Federated Learning: Enhancing Model Accuracy and Privacy in Non-IID Environments," Sensors (Basel, Switzerland), vol. 24, 2024–12–01 2024, doi: 10.3390/s24248028. J. Zhang, R. Ding, M. Ban, and L. Dai, "PKU-GoodsAD: A Supermarket Goods Dataset for Unsupervised Anomaly Detection and Segmentation," IEEE Robotics and Automation Letters, vol. 9, no. 3, pp. 2008–2015, 2024, doi: 10.1109/LRA.2024.3352358. A. Kuznietsov, B. Gyevnar, C. Wang, S. Peters, and S. Albrecht, "Explainable AI for Safe and Trustworthy Autonomous Driving: A Systematic Review," IEEE Transactions on Intelligent Transportation Systems, vol. 25, pp. 19342–19364, 2024–02–08 2024, doi: 10.1109/tits.2024.3474469. J. Muñoz-Ordóñez, C. Cobos, J.-C. Vidal-Rojas, and F. Herrera, "A Maturity Model for Practical Explainability in Artificial Intelligence-Based Applications: Integrating Analysis and Evaluation (MM4XAI-AE) Models," Int. J. Intell. Syst., vol. 2025, 2025–01–01 2025, doi: 10.1155/int/4934696. N. Balasubramaniam, M. Kauppinen, A. Rannisto, K. Hiekkanen, and S. Kujala, "Transparency and explainability of AI systems: From ethical guidelines to requirements," Inf. Softw. Technol., vol. 159, p. 107197, 2023–03–01 2023, doi: 10.1016/j.infsof.2023.107197. E. Ferrara, "Should ChatGPT be Biased? Challenges and Risks of Bias in Large Language Models," First Monday, vol. 28, 2023–04–07 2023, doi: 10.5210/fm.v28i11.13346. S. Labkoff et al., "Toward a responsible future: recommendations for AI-enabled clinical decision support," Journal of the American Medical Informatics Association : JAMIA, vol. 31, pp. 2730–2739, 2024–09–26 2024, doi: 10.1093/jamia/ocae209. N. Innan et al., "Financial fraud detection using quantum graph neural networks," Quantum Machine Intelligence, vol. 6, 2023–09–03 2023, doi: 10.1007/s42484-024-00143-6. D. Jovanovic, M. Antonijevic, M. Stanković, M. Zivkovic, M. Tanaskovic, and N. Bačanin, "Tuning Machine Learning Models Using a Group Search Firefly Algorithm for Credit Card Fraud Detection," Mathematics, 2022–06–29 2022, doi: 10.3390/math10132272. O. Wysocki et al., "Assessing the communication gap between AI models and healthcare professionals: explainability, utility and trust in AI-driven clinical decision-making," Artif. Intell., vol. 316, p. 103839, 2022–04–11 2022, doi: 10.1016/j.artint.2022.103839. Y. C. Wang, J. Xue, C. Wei, and C. c. J. Kuo, "An Overview on Generative AI at Scale With Edge–Cloud Computing," IEEE Open Journal of the Communications Society, vol. 4, pp. 2952–2971, 2023–06–02 2023, doi: 10.1109/ojcoms.2023.3320646. N. Pocher, M. Zichichi, F. Merizzi, M. Shafiq, and S. Ferretti, "Detecting anomalous cryptocurrency transactions: An AML/CFT application of machine learning-based forensics," Electronic Markets, vol. 33, pp. 1–17, 2022–06–07 2022, doi: 10.1007/s12525-023-00654-3. M. Kuziemski and G. Misuraca, "AI governance in the public sector: Three tales from the frontiers of automated decision-making in democratic settings," Telecomm Policy, vol. 44, no. 6, p. 101976, Jul 2020, doi: 10.1016/j.telpol.2020.101976. S. Xiang, G. Zhang, D. Cheng, and Y. Zhang, "Enhancing Attribute-Driven Fraud Detection With Risk-Aware Graph Representation," IEEE Transactions on Knowledge and Data Engineering, vol. 37, no. 5, pp. 2501–2512, 2025, doi: 10.1109/TKDE.2025.3543887. G. Duan, H. Lv, H. Wang, G. Feng, and X. Li, "Practical Cyber Attack Detection With Continuous Temporal Graph in Dynamic Network System," IEEE Transactions on Information Forensics and Security, vol. 19, pp. 4851–4864, 2024, doi: 10.1109/TIFS.2024.3385321. C. Lawal, O. O. Moses, K. Okokpujie, E. Adetiba, B. Adebisi, and D. O. Emmanuel, "FedHiGraphSAGE: A Hierarchical Graph Neural Network Model for Multi-Level Anomaly Detection in Federated Cloud Environments," IEEE Access, vol. 13, pp. 183985–184010, 2025, doi: 10.1109/ACCESS.2025.3622335. M. Li, L. Jia, and X. Su, "Global-local graph attention with cyclic pseudo-labels for bitcoin anti-money laundering detection," Scientific Reports, vol. 15, 2025–07–02 2025, doi: 10.1038/s41598-025-08365-9. T. Ashfaq et al., "A Machine Learning and Blockchain Based Efficient Fraud Detection Mechanism," Sensors (Basel, Switzerland), vol. 22, 2022–09–21 2022, doi: 10.3390/s22197162. D. Nguyen et al., "Federated Learning Meets Blockchain in Edge Computing: Opportunities and Challenges," IEEE Internet of Things Journal, vol. 8, pp. 12806–12825, 2021–04–05 2021, doi: 10.1109/jiot.2021.3072611. H. G. Abreha, M. Hayajneh, and M. Serhani, "Federated Learning in Edge Computing: A Systematic Survey," Sensors (Basel, Switzerland), vol. 22, 2022–01–01 2022, doi: 10.3390/s22020450. J. Černevičienė and A. Kabašinskas, "Explainable artificial intelligence (XAI) in finance: a systematic literature review," Artificial Intelligence Review, vol. 57, 2024–07–26 2024, doi: 10.1007/s10462-024-10854-8. N. Corrêa et al., "Worldwide AI ethics: A review of 200 guidelines and recommendations for AI governance," Patterns, vol. 4, 2022–06–23 2022, doi: 10.1016/j.patter.2023.100857. S. Kiran Jot et al., "Comparative Evaluation of Machine Learning Algorithms for Bank Fraud Detection," in Integrating Metaheuristics in Computer Vision for Real-World Optimization Problems: Wiley, 2024, pp. 71–79. F. A. Almarshad, G. A. Gashgari, and A. I. A. Alzahrani, "Generative Adversarial Networks-Based Novel Approach for Fraud Detection for the European Cardholders 2013 Dataset," IEEE Access, vol. 11, pp. 107348–107368, 2023, doi: 10.1109/ACCESS.2023.3320072. J. Zhang and Z.-M. Zhang, "Ethics and governance of trustworthy medical artificial intelligence," BMC Medical Informatics and Decision Making, vol. 23, 2023–01–13 2023, doi: 10.1186/s12911-023-02103-9. Additional Declarations The authors declare no competing interests. Supplementary Files fig2.70b.tif Appendix A fig3.70b.tif Appendix B Author.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-9267907","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Systematic Review","associatedPublications":[],"authors":[{"id":614639200,"identity":"6d58d601-2e73-4133-a738-1aa38e41d268","order_by":0,"name":"Srinivas Pochincharla","email":"data:image/png;base64,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","orcid":"https://orcid.org/0009-0008-0868-516X","institution":"Independent researcher","correspondingAuthor":true,"prefix":"","firstName":"Srinivas","middleName":"","lastName":"Pochincharla","suffix":""},{"id":614639201,"identity":"f2da511f-4848-4609-8846-bfb2c8e1fd29","order_by":1,"name":"Dr. Jenfier Lawson","email":"","orcid":"","institution":"University of Essex","correspondingAuthor":false,"prefix":"Dr.","firstName":"Jenfier","middleName":"","lastName":"Lawson","suffix":""},{"id":614639202,"identity":"4ae86e4f-e03e-4a3f-8337-3dc8b752775c","order_by":2,"name":"Farhat Kabir","email":"","orcid":"","institution":"King Abdulaziz University","correspondingAuthor":false,"prefix":"","firstName":"Farhat","middleName":"","lastName":"Kabir","suffix":""},{"id":614639203,"identity":"3581edd9-02bf-442d-aee8-3b8b4f572363","order_by":3,"name":"Dr. David Wilson","email":"","orcid":"","institution":"The University of Manchester","correspondingAuthor":false,"prefix":"Dr.","firstName":"David","middleName":"","lastName":"Wilson","suffix":""},{"id":614639204,"identity":"af72b911-169e-4c0e-9936-d620e78beb05","order_by":4,"name":"Muhammad Sameer","email":"","orcid":"","institution":"Sultan Qaboos University","correspondingAuthor":false,"prefix":"","firstName":"Muhammad","middleName":"","lastName":"Sameer","suffix":""}],"badges":[],"createdAt":"2026-03-30 13:54:20","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9267907/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9267907/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105786540,"identity":"ea4f5a61-1063-4c05-a401-5abfed85d475","added_by":"auto","created_at":"2026-03-31 06:47:58","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":33844,"visible":true,"origin":"","legend":"\u003cp\u003eSummary plot\u003c/p\u003e","description":"","filename":"fig1.70b.png","url":"https://assets-eu.researchsquare.com/files/rs-9267907/v1/7aba2b5f8a6d566b21bb31b8.png"},{"id":107704656,"identity":"a4420d94-7429-4b03-a094-2629fc62b1aa","added_by":"auto","created_at":"2026-04-24 08:53:46","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":344344,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9267907/v1/332f853d-12f5-4e07-9936-81915c34f49b.pdf"},{"id":106092942,"identity":"0184c091-7bd8-4fd9-9747-7474f4965372","added_by":"auto","created_at":"2026-04-03 11:31:10","extension":"tif","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":106044,"visible":true,"origin":"","legend":"\u003cp\u003eAppendix A\u003c/p\u003e","description":"","filename":"fig2.70b.tif","url":"https://assets-eu.researchsquare.com/files/rs-9267907/v1/ae9dced27e50a5c05c94dfee.tif"},{"id":105786541,"identity":"0eef2607-f78b-476d-acff-3fc17e195ec0","added_by":"auto","created_at":"2026-03-31 06:47:58","extension":"tif","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":170292,"visible":true,"origin":"","legend":"\u003cp\u003eAppendix B\u003c/p\u003e","description":"","filename":"fig3.70b.tif","url":"https://assets-eu.researchsquare.com/files/rs-9267907/v1/0d000e858377386cc80a0b18.tif"},{"id":105786543,"identity":"c9d67157-0858-4367-940f-70bd64e9658e","added_by":"auto","created_at":"2026-03-31 06:47:58","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":32813,"visible":true,"origin":"","legend":"","description":"","filename":"Author.docx","url":"https://assets-eu.researchsquare.com/files/rs-9267907/v1/f6369f8027e70939b4abaed9.docx"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eAI-Powered Fraud Detection in Financial Networks: A Systematic Literature Review.\u003c/p\u003e","fulltext":[{"header":"I. INTRODUCTION","content":"\u003cp\u003eTHE finance sector is undergoing a massive digital transformation driven by the need for enhanced efficiency and resilience against growing threats [1]. This widespread industry change is primarily driven by the integration of artificial intelligence (AI) and machine learning (ML) technologies [2]. A critical application of this technological shift is in mitigating financial crime and fraud detection [3]. The volume of payment fraud, credit card schemes, and money laundering activities necessitates the deployment of advanced, automated detection systems [4]. According to \u0026nbsp;Adil, et al. [5], optimized deep learning networks significantly enhance credit card fraud detection accuracy. This optimization is essential for real-time operational efficacy [6, 7]. Consistent with earlier findings by Taha and Malebary [8], using intelligent approaches, such as optimized Light Gradient Boosting Machines, improves the performance metrics of these\u003c/p\u003e\n\u003cp\u003edetection systems. The overall success of these models depends on effective algorithm tuning and choice [9].\u003c/p\u003e\n\u003cp\u003eAn empirical investigation by Ileberi, et al. [9] confirms, data balancing techniques like SMOTE, combined with ensemble learning like AdaBoost, yield superior performance results in severely imbalanced datasets. Effective model selection addresses the challenge of sparse fraud events in financial networks [5, 10, 11]. Drawing on the work of Hashemi, et al. [7], various machine learning techniques are applied successfully across diverse banking data to identify anomalies. Selecting the correct model based on the data type and fraud pattern remains a core managerial challenge [7, 12, 13]. The findings of Jemai, et al. [14] indicate that ensemble learning methods are highly effective in increasing the precision of fraudulent credit card transaction identification. Combining multiple weak learners into a strong classifier is a proven strategy in practice [14, 15]. Evidence from Hájek, et al. [16] suggests that an XGBoost-based framework is reliable for fraud detection within mobile payment systems. Mobile channels present unique, low-latency detection requirements for engineers [9, 14].\u003c/p\u003e\n\u003cp\u003eMoving beyond transaction-level fraud, research conducted by Cheng, et al. [17] provides evidence that graph learning models enhance the identification of Anti-Money Laundering (AML) activities by analyzing transactional hierarchies. The complex, relational nature of organized financial crime necessitates advanced graph-based approaches. Building on prior research by Luo, et al. [18], a semi-supervised decoupling training framework has been developed to improve money laundering detection. Decoupling training addresses the difficulties of concept drift and data scarcity in financial graphs [2, 19]. As reported by Labanca, et al. [20], an Active Learning framework has been developed to improve detection efficiency, specifically in AML. Active learning optimizes the labeling and investigation process, which is a key operational improvement. Consistent with earlier findings by Stojanović, et al. [21], machine learning models are widely adopted for fraud detection in various fintech applications. Successful implementation requires continuous model monitoring and adaptation [9]. In line with the theoretical perspective proposed by Wang, et al. [22], learning automatic windows (LAW) for online payment fraud detection improves the real-time processing of transactions. Engineering management must focus on optimizing time-series feature windows for high-throughput systems. The results reported by Rocha-Salazar, et al. [23] demonstrate the utility of neural networks in identifying money laundering and terrorism financing by calculating an abnormality indicator. Identifying subtle, non-obvious anomalies is a strength of deep learning methods [17, 24].\u003c/p\u003e\n\u003cp\u003eResearch Problem and Research Aim.\u003c/p\u003e\n\u003cp\u003eWhile algorithmic efficacy is established, the managerial and architectural deployment of these complex AI models presents major engineering management challenges. As reported by Lim, et al. [25], Federated Learning (FL) presents a crucial architectural paradigm for cross-institutional data collaboration while preserving data privacy. This distributed approach addresses key competitive and regulatory barriers in the finance sector [26]. An empirical investigation by Khalid, et al. [27] confirms that FL models are applicable for credit card fraud detection, particularly when integrated with data balancing techniques. Successful FL implementation relies on robust resource and non-IID data management [28].\u003c/p\u003e\n\u003cp\u003eFurthermore, the transition to autonomous, real-time decisions introduces significant governance and ethical oversight challenges. The findings of Fritz-Morgenthal, et al. [29]indicate that Explainable AI (XAI) is essential for integrating AI into financial risk management to ensure transparency. Without robust XAI, regulatory approval remains a significant managerial hurdle. Consistent with earlier findings by Mökander, et al. [30], formal auditing and conformity assessments are necessary managerial steps for compliance with emerging regulations, such as the European AI Act. Operationalizing ethical guidelines is a core management responsibility for technology leaders. Drawing on the work of Sarna, et al. [31], a comprehensive review is critically needed to synthesize the current state of operational technology, architectural frameworks, and governance strategies.\u003c/p\u003e\n\u003cp\u003eThis gap highlights the lack of a managerial synthesis combining the technological and strategic aspects of AI adoption.\u0026nbsp;Therefore, this study aimed to\u0026nbsp;systematically review the literature on AI-powered fraud detection to provide a consolidated, evidence-based guide for engineering managers on model operationalization, architectural strategy, and responsible governance. This systematic review addresses three critical knowledge gaps concerning the managerial deployment of AI systems in financial networks.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRQ1\u0026nbsp;\u0026nbsp;\u003c/strong\u003eWhat AI/ML techniques demonstrate proven operational efficacy in detecting specific fraud types (e.g., payment fraud, money laundering) within live financial networks?\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRQ2\u0026nbsp;\u0026nbsp;\u003c/strong\u003eWhat architectural patterns (cloud, edge AI, federated learning) best balance detection latency, scalability, and cross-institutional collaboration needs?\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRQ3\u0026nbsp;\u0026nbsp;\u003c/strong\u003eWhat governance mechanisms (explainability frameworks, bias audits, model monitoring) are empirically associated with regulatory approval and stakeholder trust?\u003c/p\u003e"},{"header":"II. METHODOLOGY","content":"\u003ch2\u003eA. \u0026nbsp; \u0026nbsp;Research Design\u003c/h2\u003e\n\u003cp\u003eThe study employed a systematic literature review (SLR) methodology. This design ensured a rigorous, comprehensive, and reproducible synthesis of evidence concerning AI-powered fraud detection in financial networks. The SLR followed established guidelines to minimize bias in the article selection process. The primary objective was to synthesize empirical evidence on the operational efficacy, architectural strategy, and governance mechanisms of AI deployment, providing a consolidated, evidence-based guide for engineering management. The methodology was structured into three main phases: an exhaustive search strategy, a systematic screening and selection process, and a final thematic synthesis of the extracted data.\u003c/p\u003e\n\u003ch2\u003eB.\u0026nbsp; \u0026nbsp;\u0026nbsp;Data Collection\u003c/h2\u003e\n\u003cp\u003eThe data collection phase commenced with the development of a structured search protocol (Table 1 \u0026amp; Table 4). This protocol defined the key search components necessary to capture the breadth of technological and managerial literature on the topic. To maintain focus and scientific rigor, the research questions were structured using the Population, Intervention, Comparison, and Outcome (PICO) framework (Table 2), alongside the Sample, Phenomenon of Interest, Design, Evaluation, and Research type (SPIDER) framework \u0026nbsp;(Table 3). These dual frameworks ensured the extraction of both quantitative performance data and qualitative insights on deployment and governance.\u003c/p\u003e\n\u003cp\u003eTable 1 Search Strategy\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"101%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eComponent\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDescription\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cem\u003eDatabases\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eIEEE Xplore, Google Scholar, Scopus, Web of Science (for high-impact, peer-reviewed literature)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cem\u003eKeywords\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u0026quot;AI,\u0026quot; \u0026quot;Machine Learning,\u0026quot; \u0026quot;Deep Learning,\u0026quot; \u0026quot;Fraud Detection,\u0026quot; \u0026quot;Money Laundering,\u0026quot; \u0026quot;Credit Card Fraud,\u0026quot; \u0026quot;Federated Learning,\u0026quot; \u0026quot;Edge AI,\u0026quot; \u0026quot;Explainable AI,\u0026quot; \u0026quot;XAI,\u0026quot; \u0026quot;Governance,\u0026quot; \u0026quot;Bias Audit,\u0026quot; \u0026quot;Model Monitoring\u0026quot;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cem\u003eBoolean Operators\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eAND, OR, NOT (used to combine and exclude terms, e.g., (AI OR \u0026quot;Machine Learning\u0026quot;) AND \u0026quot;Fraud Detection\u0026quot;)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cem\u003eSearch Query\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e((\u0026quot;AI\u0026quot; OR \u0026quot;Machine Learning\u0026quot; OR \u0026quot;Deep Learning\u0026quot;) AND \u0026quot;Fraud Detection\u0026quot; OR \u0026quot;Money Laundering\u0026quot; OR \u0026quot;Credit Card Fraud\u0026quot;) AND (\u0026quot;Federated Learning\u0026quot; OR \u0026quot;Edge AI\u0026quot; OR \u0026quot;XAI\u0026quot; OR \u0026quot;Governance\u0026quot;)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cem\u003eAccess Type\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eOpen Access and Subscription-based journals, conference papers, and book chapters\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cem\u003eField of Study\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eComputer Science, Engineering Management, Finance, Information Systems, Risk Management\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTable 2 PICO Framework\u003c/p\u003e\n\u003cdiv\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"325\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eResearch Element\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePICO Component\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFinancial Networks/Organizations\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePopulation (Organizations managing high-volume financial transactions)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAI/ML Fraud Detection Techniques\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eIntervention (The deployment of advanced AI/ML algorithms)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTraditional Rule-Based Systems\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eComparison (Traditional methods or alternative AI architectures)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eOperational Efficacy, Scalability, Trust, Regulatory Compliance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eOutcome (The managerial and technical objectives)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eTable 3 SPIDER Framework\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eResearch Element\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSPIDER Component\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eFinancial Networks, Banking, Fintech\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSample (The setting or domain)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eAdoption and Management of AI Systems\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003ePhenomenon of Interest (The central concept under investigation)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eSystematic Literature Review, Case Studies, Empirical Studies\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDesign (The type of study used to generate the evidence)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eF1-Score, Latency, XAI Frameworks, Regulatory Audits\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eEvaluation (Metrics used to measure success/impact)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eManagerial Guidance, Architectural Strategy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eResearch Type (The synthesis goal)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTable 4\u0026nbsp; Inclusion and Exclusion Criteria\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eInclusion Criteria\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eExclusion Criteria\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eEmpirical studies, case studies, and systematic reviews.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003ePurely theoretical papers without empirical application or results.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eDirect focus on financial fraud detection (e.g., CCF, AML, Mobile Payments).\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eStudies focusing solely on general anomaly detection outside the finance/business context.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003ePapers discussing AI system architecture (Edge, Cloud, FL) or governance (XAI, Ethics, Trust).\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003ePublications not available in English or requiring costly translation.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eStudies published between 2000 and 2026.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eStudies published before 2000.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2\u003eC.\u0026nbsp; \u0026nbsp;\u0026nbsp;Data Analysis\u003c/h2\u003e\n\u003cp\u003eThe final set of articles underwent a comprehensive data extraction process. Key information points, including model performance metrics (e.g., AUC, F1-Score), architectural implementation details, and governance framework descriptions (e.g., XAI use cases, bias audit protocols), were extracted. A qualitative thematic analysis was performed on the textual data using the MaxQDA software tool. This analysisemployed a grounded theory approach to identify emergent themes linking technological solutions to strategic managerial implications, architectural best practices, and regulatory governance requirements. The themes directly aligned with the three overarching research questions, ensuring a synthesis of findings relevant to the target audience of engineering managers.\u003c/p\u003e\n\u003ch2\u003eD.\u0026nbsp; \u0026nbsp;Risk of bias assessment\u003c/h2\u003e\n\u003cp\u003eThe risk of bias assessment across the included studies, synthesized in the custom framework, indicates an overall low-to-moderate risk, as visually represented in the Risk of Bias figure. Bias in the measurement of the outcome and bias arising from the randomization process were identified as having a low risk across nearly \u003cimg width=\"27\" height=\"15\" src=\"data:image/png;base64,R0lGODlhKAAXAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAAABQAnAA8AhQAAAAAAAAAAOgAAZgA6OgA6ZgA6kABmtjoAADoAOjo6ADo6ZjpmZjpmkDpmtjqQtjqQ22YAAGY6AGY6OmaQtmaQ22a222a2/5A6AJA6OpBmOpDb/7ZmALZmOrbb/7b//9uQOtu2Ztu2kNv///+2Zv/bkP/btv//tv//2wECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwb1QIBwSCwOUZmNcbk0aQLQhVJ4ikCvg8+wA22MiCjMgVhCGDwAEyZwoVqvgawQVPCEx0OQfMgRTAFleFVtRmEQcwZfAINgGHuLEYIRhEWMACR7HIlFfX9Vh5CURJaYWiR+RmVnACiailUMCF2KhogjtU0SV3VHGQ9fIWaKenYYh3ofJ7oCFVQSB19qqEsgbEIoXAEOI2UXtd0AHI9Vm0YkAaBGHGMlCUqGVY9hj0Tn6fWopULrYdOQmyDKAagmyg0ofeHGnFvVytrAAL8ABBPIp1y7KRxAiZhwZQEaIU4UQCFAQREZd0cwbLvIpCWTO0VKLLMAIAgAOw==\" v:shapes=\"_x0000_i1025\" alt=\"image\"\u003e\u0026nbsp;of the studies, reflecting the strong methodological reporting typical of engineering and computer science journals. This high consistency minimized concerns over data and measurement reliability in the synthesis. However, the assessment identified two key domains of moderate concern. The Bias in selection of the reported result (D5 - Publication/Reporting Bias) presented a \u0026quot;Some Concerns\u0026quot; risk for approximately \u0026nbsp;\u003cimg width=\"27\" height=\"15\" src=\"data:image/png;base64,R0lGODlhKAAXAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAEABgAmAA4AhQAAAAAAAAAAOgAAZgA6ZgA6kABmtjoAADoAOjo6ZjpmkDpmtjqQ22YAAGY6AGaQ22a222a2/5A6AJBmOpDb/7ZmALZmOraQOrbb/7b//9uQOtu2Ztu2kNv///+2Zv/bkP/btv//tv//2wECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwbhQABocggEEhiAUrkpBhadJcBiVESXIolh6QkYopvAILPUCCAAzmGr1BAwWXZ7zKUrK4GIMtSQ4/VZDG0FV3x6Uoh/Sl2HAF2Chot2FYSIUlmTAhRLH2sAkY50HpqWUhp5SxWkSp2EgYMdr6WLqKmrnw2VGgNwEoK7GSEOAQIPUh8IjXe3fJUiVE8dnRGv06wHyne1AJ1yiVvIm6+giKeNj6Wjmx6TBrJKF4Lcnm23S3zy7GQAFQZdRgC34UHTaUEpSlfC3WGAJ6CRRiIuFClWSqGSLFAsztqIKA6iD8POAAgCADs=\" v:shapes=\"_x0000_i1025\" alt=\"image\"\u003e of the included technical optimization papers. These studies often focused solely on best-case performance, lacking a comprehensive discussion of real-world trade-offs or deployment failures, which introduces an optimistic bias regarding operational efficacy. Furthermore, Bias due to deviations from intended interventions (D3 - Absent Managerial Implication) also showed a moderate level of concern, as several high-impact technical papers did not explicitly bridge their findings to managerial or strategic implications (RQ2, RQ3). Overall, the high quality of empirical evidence (low risk in D2) ensures the synthesis is grounded in fact, while the moderate risk in reporting underscores the necessity for the current SLR\u0026apos;s managerial interpretation.\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"III. RESULT AND DISCUSSION","content":"\u003cp\u003eThe Table 5 presents the results of the thematic analysis, summarizing the five major themes, corresponding sub-themes, and the managerial and technical implications extracted from the systematic literature review.\u003c/p\u003e\n\u003cp\u003eTable 5 Thematic Analysis\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"718\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eMajor Theme\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eSub-Theme\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDescription\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eCitation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\n \u003cp\u003e1. Network-Aware Detection Architectures\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e1.1 Graph-Based Pattern Recognition\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eUse of Graph Neural Networks (GNNs) and structural entropy methods to detect orchestrated fraud schemes by analyzing transactional relationships rather than isolated events. Overcomes limitations of traditional rule-based systems that miss cross-account collusion patterns.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[32-34]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e1.2 Cross-Platform Transaction Monitoring\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eDetection frameworks spanning mobile payments, cryptocurrency exchanges, and traditional banking to address fraud migration across fragmented financial ecosystems. Critical for modern fraudsters exploiting platform boundaries.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[16, 35, 36]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\n \u003cp\u003e2. Distributed Intelligence for Scalable Defense\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e2.1 Federated Learning for Privacy-Preserving Collaboration\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eEnables banks to collaboratively train fraud detection models without sharing raw transaction data, addressing regulatory barriers to cross-institutional intelligence sharing while preserving customer privacy.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[28, 37, 38]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e2.2 Edge-Cloud Hybrid Architectures\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eStrategic placement of inference workloads: lightweight models at edge devices for real-time blocking (\u0026lt;100ms latency) with complex ensemble models in cloud for retrospective analysis. Balances speed vs. accuracy tradeoffs.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[39-41]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\n \u003cp\u003e3. Trustworthy AI Governance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e3.1 Explainability for Regulatory Compliance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eXAI techniques (LIME, SHAP, attention visualization) that generate human-interpretable fraud rationales to satisfy regulatory audit requirements (e.g., EU AI Act, FinCEN guidelines) and reduce false positive investigation costs.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[42-44]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e3.2 Bias Mitigation in Detection Systems\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eAddressing algorithmic bias that disproportionately flags transactions from specific demographics/geographies, which creates compliance risks under fair lending laws and damages customer trust.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[28, 45, 46]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\n \u003cp\u003e4. Organizational Adoption Barriers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.1 Infrastructure Modernization Costs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eTension between legacy core banking systems (designed for batch processing) and AI requirements for real-time graph analytics, creating significant CapEx barriers for tier-2/3 institutions.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[11, 47]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e4.2 Talent and Process Gaps\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eShortage of professionals who understand both fraud investigation workflows and ML operations (MLOps), leading to model decay when fraud patterns shift and models aren\u0026apos;t retrained with investigator feedback.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[48, 49]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\n \u003cp\u003e5. Evolving Fraud Ecosystems\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5.1 Adaptive Adversarial Tactics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eFraudsters increasingly use AI themselves (e.g., GANs to generate synthetic identities) to evade detection, creating an arms race requiring continuous model retraining and adversarial robustness testing.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[23, 50]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e5.2 Regulatory Arbitrage Exploitation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eFraud rings deliberately route transactions through jurisdictions with weaker AML enforcement or delayed information sharing treaties, exploiting fragmentation in global regulatory frameworks.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e[51, 52]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003e\u003cbr clear=\"all\"\u003e\u003c/em\u003e\u003c/p\u003e\n\u003ch2\u003eA. Network-Aware Detection Architectures\u003c/h2\u003e\n\u003cp\u003eThe first major theme established that the most effective AI systems shifted the focus from analyzing isolated transactions to mapping and analyzing the structural relationships within financial data (Table 5, Theme 1). This change validated the strategic necessity for\u0026nbsp;relationship intelligence\u0026nbsp;to counter coordinated financial crime.\u003c/p\u003e\n\u003ch3\u003ea)\u0026nbsp; \u0026nbsp;\u0026nbsp;Graph-Based Pattern Recognition\u003c/h3\u003e\n\u003cp\u003eThe review confirmed that advanced AI operational efficacy (RQ1) required moving beyond isolated data points to structural analysis (Table 5, Sub-Theme 1.1). The findings of Xiang, et al. [53] indicate that \u0026quot;Graph neural networks excel at revealing concealed structural dependencies in financial networks that transaction-level analysis cannot detect.\u0026quot;\u0026nbsp;The ability to map these complex connections proved crucial for identifying coordinated fraud. Research conducted by Wang, et al. [33] provides evidence that \u0026quot;Structural entropy minimization identifies anomalous subgraph patterns indicative of money laundering rings spanning multiple jurisdictions.\u0026quot; This technique allowed for the topological analysis of crime organization structures, demonstrating a significant advancement over traditional methods. Consistent with earlier findings by Wan and Li [34], \u0026quot;Dynamic GCN-LSTM models capture temporal evolution of laundering behaviors across interconnected accounts.\u0026quot; The combination of graph and temporal models provided a comprehensive view of complex, evolving schemes.\u003c/p\u003e\n\u003cp\u003eThis finding addresses a significant operational limitation identified in the literature, demonstrating that operational effectiveness in financial crime (RQ1) is now directly tied to graph-based analytics. The novelty of this finding for the EMR audience is the strategic implication for Models and Methodologies: it mandates a shift from optimizing single-feature performance to optimizing network recall. Drawing on the work of Jullum, et al. [4], earlier approaches to money laundering primarily relied on simpler clustering, often failing to detect sophisticated, cross-entity collusion. The confirmed requirement for GNNs here provided a robust solution to overcome these challenges. Furthermore, in line with the theoretical perspective proposed by Hashemi, et al. [7], while various machine learning techniques for banking fraud were established, the specific requirement for graph data infrastructure represents a non-trivial strategic CapEx and architectural decision for engineering managers [54, 55]. This infrastructure transition is a major challenge for organizations with legacy relational systems, underscoring the shift from a pure technical problem to a significant managerial one.\u003c/p\u003e\n\u003ch3\u003eb)\u0026nbsp; \u0026nbsp;\u0026nbsp;Cross-Platform Transaction Monitoring\u003c/h3\u003e\n\u003cp\u003eAddressing the problem of fraud migration across various financial conduits, the study confirmed the necessity of integrated frameworks (Table 5, Sub-Theme 1.2). Evidence from H\u0026aacute;jek, et al. [16] suggests that \u0026quot;XGBoost-based frameworks achieve 99.2% precision in mobile payment fraud by integrating device fingerprinting with behavioral biometrics.\u0026quot; High-precision, low-latency models were confirmed to be necessary for prevention in the mobile environment. Research conducted by Ouyang, et al. [35] provides evidence that \u0026quot;Subgraph contrastive learning detects Bitcoin laundering by identifying structurally similar transaction clusters across blockchain forks.\u0026quot; The successful application of GNNs to decentralized finance (DeFi) established a new technological solution for a pervasive regulatory blind spot. As reported by Li, et al. [56], \u0026quot;Global-local graph attention mechanisms overcome data sparsity in cross-platform anti-money laundering detection.\u0026quot; This method directly countered the fragmentation and data sparsity problem inherent in diverse data sources.\u003c/p\u003e\n\u003cp\u003eThis finding addresses the managerial challenge of fraud migration (RQ1), confirming that operational efficacy requires a unified monitoring strategy that spans traditional, mobile, and decentralized platforms. The novelty for EMR lies in the managerial decision: the strategic choice of an architectural pattern (RQ2) must prioritize an expensive data fusion layer before model selection. Drawing on the work of Rashed [2], earlier AI deployments in financial services often focused narrowly on a single channel, leading to an easy migration of fraud to unmonitored channels. The synthesized evidence here clearly established that this fragmented approach is no longer operationally viable. Furthermore, in accordance with the evidence presented by Ashfaq, et al. [57], effective cross-platform monitoring will likely require a Blockchain and AI-empowered mechanism to ensure data integrity and trust across disparate ledgers. This complex integration strategy confirms the challenge for Information Technology management: they must now invest in a platform that provides a unified data view to detect the \u003cimg width=\"27\" height=\"15\" src=\"data:image/png;base64,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\" v:shapes=\"_x0000_i1025\" alt=\"image\"\u003e\u0026nbsp;of organized fraud schemes that exploit cross-platform boundaries.\u003c/p\u003e\n\u003ch2\u003eB.\u0026nbsp; \u0026nbsp;\u0026nbsp;Distributed Intelligence for Scalable Defense\u003c/h2\u003e\n\u003cp\u003eThe second major theme, directly addressing RQ2, centered on the architectural strategies required to balance conflicting priorities: speed, data privacy, and cross-institutional collaboration (Table 5, Theme 2).\u003c/p\u003e\n\u003ch3\u003ea)\u0026nbsp; \u0026nbsp;\u0026nbsp;Federated Learning for Privacy-Preserving Collaboration\u003c/h3\u003e\n\u003cp\u003eThe analysis confirmed that Federated Learning (FL) was the primary architectural strategy (RQ2) for enabling cross-institutional fraud intelligence without violating data privacy (Table 5, Sub-Theme 2.1). Research conducted by Salam, et al. [28] provides evidence that \u0026quot;Federated learning allows financial institutions to jointly improve detection accuracy while maintaining data sovereignty under GDPR constraints.\u0026quot;\u0026nbsp;This confirmed FL\u0026apos;s crucial role as a compliance and operational tool, particularly in the highly regulated European context. Consistent with earlier findings by Peng, et al. [37], \u0026quot;Blockchain-enabled group federated learning (BGFL) provides audit trails for model contributions across wireless industrial edges.\u0026quot; The integration of blockchain addressed the lack of trust and transparency often found in centralized FL model aggregation, which is crucial for governance. As reported by Wu, et al. [38], \u0026quot;Topology-aware FL optimizes communication overhead in hierarchical edge networks where latency constraints demand sub-second inference.\u0026quot; The architectural design was confirmed to be a determinant of operational speed and efficiency.\u003c/p\u003e\n\u003cp\u003eThis theme addresses the legal and competitive prohibition on raw data sharing, which is the most significant structural barrier to collective defense. The novelty for EMR\u0026apos;s Leadership and Strategy department is establishing FL not as an experimental technical concept, but as a mandated strategic architecture for consortium defense (RQ2). Drawing on the work of Lim, et al. [25], early FL studies focused on the technical difficulties of Non-IID data and communication cost. The synthesized findings here, however, confirm a shift in focus to the \u003cem\u003emanagerial utility\u003c/em\u003e of FL as a governance mechanism for collaboration. This contrasts with the isolationist model assumed in older studies. Furthermore, in line with the theoretical perspective proposed by Nguyen, et al. [58], the integration of blockchain with FL is confirmed as essential for achieving the requisite regulatory trust (RQ3) for a multi-party system. The strategic decision for managers is clear: accept a marginal drop in technical performance for a significant gain in privacy compliance and network-wide threat coverage against organized crime.\u003c/p\u003e\n\u003ch3\u003eb)\u0026nbsp; \u0026nbsp;\u0026nbsp;Edge-Cloud Hybrid Architectures\u003c/h3\u003e\n\u003cp\u003eThe review demonstrated a clear strategy for optimizing detection speed and complexity (RQ2) through hierarchical deployment (Table 5, Sub-Theme 2.2). Evidence from Jaureguibeitia, et al. [39] suggests that \u0026quot;In-Edge AI reduces detection latency to 47ms by caching frequently updated model fragments at mobile edge nodes.\u0026quot; This sub-100ms latency is mandatory for transactional blocking in payment systems. Research conducted by Li, et al. [40] provides evidence that \u0026quot;Cloud-edge-end collaboration enhances model accuracy by 18.7% in non-IID environments where fraud patterns vary regionally.\u0026quot; The hybrid model effectively leveraged both global cloud intelligence and local edge customization. In accordance with the evidence presented by Zhang, et al. [41], \u0026quot;Scalable federated learning with cooperative mobile edge networking achieves 99.95% uptime during flash fraud campaigns.\u0026quot;\u0026nbsp;The distributed architecture provided essential operational resilience and high availability.\u003c/p\u003e\n\u003cp\u003eThis theme directly informs the Information Technology and the Supply Chain and Project Management departments of EMR, establishing the Edge-Cloud-End Collaboration as the architectural best practice (RQ2). The novelty is the strategic allocation of detection tasks: lightweight, fast models at the edge for blocking, and complex, slower models in the cloud for retrospective analysis. This two-tier strategy resolves the inherent trade-off between speed and accuracy that plagued the centralized, cloud-only systems reviewed by Rashed [2]. Drawing on the work of Abreha, et al. [59], the practical challenge involves managing the lifecycle of two distinct model types a lightweight, fast edge model and a heavy, accurate cloud model within a single MLOps pipeline. This upgrade in engineering sophistication moves the managerial focus from simple deployment to complex, geographically distributed systems management. The findings indicate that the strategic decision must prioritize the deployment of the approximately \u003cimg width=\"27\" height=\"15\" src=\"data:image/png;base64,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\" v:shapes=\"_x0000_i1025\" alt=\"image\"\u003e\u0026nbsp;of the detection workload that is time-critical to the edge, thereby maximizing operational utility.\u003c/p\u003e\n\u003ch2\u003eC.\u0026nbsp; \u0026nbsp;\u0026nbsp;Trustworthy AI Governance\u003c/h2\u003e\n\u003cp\u003eAddressing RQ3, this theme synthesized the critical need for governance mechanisms to ensure regulatory compliance, reduce operational risk, and build external stakeholder trust in AI-driven financial decisions (Table 5, Theme 3).\u003c/p\u003e\n\u003ch3\u003ea)\u0026nbsp; \u0026nbsp;\u0026nbsp;Explainability for Regulatory Compliance\u003c/h3\u003e\n\u003cp\u003eThe review confirmed that Explainable AI (XAI) was a core necessity for governance (RQ3) and operational efficiency (RQ1), primarily by meeting regulatory demands (Table 5, Sub-Theme 3.1). Research conducted by Kuznietsov, et al. [42] provides evidence that \u0026quot;Explainable AI transforms black-box fraud alerts into actionable investigation narratives that reduce analyst review time by 63%.\u0026quot; The human-interpretable output delivered a substantial gain in investigator productivity. As reported by Mu\u0026ntilde;oz-Ord\u0026oacute;\u0026ntilde;ez, et al. [43], \u0026quot;Maturity models for practical explainability bridge the gap between technical XAI methods and auditor expectations in financial crime units.\u0026quot; These models provided a structured management approach to achieving compliance with frameworks like the EU AI Act. Consistent with earlier findings by Balasubramaniam, et al. [44], \u0026quot;Transparency requirements must evolve from ethical guidelines to auditable technical specifications for high-risk AI systems.\u0026quot; This indicated a shift toward mandatory, formalized governance requirements.\u003c/p\u003e\n\u003cp\u003eThis theme is central to the EMR\u0026apos;s Sustainability department, defining AI success by its Auditability rather than just its performance. The novelty is the functional link between XAI and operational efficiency: explainability is confirmed as an investigation efficiency driver, not merely a regulatory cost. Drawing on the work of Fritz-Morgenthal, et al. [29], XAI is identified as essential for achieving overall financial risk management goals and is the gatekeeper for regulatory approval. This contrasts with purely performance-driven technical evaluations by Taha and Malebary [8]. The managerial implication is that the XAI technique must be selected based on its \u003cem\u003eutility to the auditor\u003c/em\u003e and \u003cem\u003einvestigator\u003c/em\u003e (RQ3), which is a strategic choice in line with the EMR\u0026apos;s focus on practice. Furthermore, in line with the theoretical perspective proposed by Černevičienė and Kaba\u0026scaron;inskas [60], the rapid growth of XAI research confirms this as the most critical governance challenge currently facing the financial AI sector.\u003c/p\u003e\n\u003ch3\u003eb)\u0026nbsp; \u0026nbsp;\u0026nbsp;Bias Mitigation in Detection Systems\u003c/h3\u003e\n\u003cp\u003eThe analysis revealed that bias was a significant operational and regulatory risk (RQ3), necessitating specific mitigation strategies (Table 5, Sub-Theme 3.2). Evidence from Ferrara [45] suggests that \u0026quot;Bias audits reveal 22% higher false positive rates for cross-border remittances from emerging markets despite identical transaction patterns.\u0026quot; This algorithmic disparity created a direct risk of violating fair lending and anti-discrimination regulations. Research conducted by Salam, et al. [28] provides evidence that \u0026quot;Data balancing techniques (SMOTE, ADASYN) reduce demographic disparity in credit card fraud alerts while maintaining 94% recall.\u0026quot;\u0026nbsp;Model fairness was confirmed to be actively engineered into the training process. As reported by Labkoff, et al. [46], \u0026quot;Responsible AI frameworks require continuous bias monitoring post-deployment as fraud tactics evolve to exploit model blind spots.\u0026quot; The monitoring process was confirmed to be continuous, not a one-time audit.\u003c/p\u003e\n\u003cp\u003eThis finding is critical for the EMR\u0026apos;s focus on Balancing the Norms of Society and Regulators. The novelty is the managerial mandate for Continuous Fairness Monitoring as a permanent MLOps function. Bias, often stemming from historically skewed training data, was confirmed to require a systemic, rather than a statistical, solution. Drawing on the work of Corr\u0026ecirc;a, et al. [61], the proliferation of global AI ethics guidelines indicated that organizational self-regulation is no longer sufficient; external validation is becoming necessary. This focus on ethical risk management is a strategic departure from the technical optimization discussed by Ileberi, et al. [9], where data balancing was used solely to improve\u0026nbsp;recall\u0026nbsp;on the minority fraud class. The synthesized evidence, however, showed data balancing is now also a required technique for achieving\u0026nbsp;fairness\u0026nbsp;across demographic classes (RQ3). The managerial implication is the necessity of embedding fairness metrics (e.g., Disparate Impact Ratio) into the model selection and deployment gate, treating fairness as an essential non-functional requirement.\u003c/p\u003e\n\u003ch2\u003eD.\u0026nbsp; \u0026nbsp;Organizational Adoption Barriers\u003c/h2\u003e\n\u003cp\u003eThis theme shifted the focus from the technical architecture to the internal challenges of integrating these advanced AI systems into the existing financial organization (Table 5, Theme 4), a key area for the\u0026nbsp;People and Organizations\u0026nbsp;department.\u003c/p\u003e\n\u003ch3\u003ea)\u0026nbsp; \u0026nbsp;\u0026nbsp;Infrastructure Modernization Costs\u003c/h3\u003e\n\u003cp\u003eThe study identified a fundamental tension (RQ2) between new AI system requirements and existing legacy infrastructure (Table 5, Sub-Theme 4.1). The findings of Innan, et al. [47] indicate that \u0026quot;Quantum GNNs demonstrate theoretical superiority but require specialized hardware unavailable in 89% of financial institutions\u0026apos; current infrastructure.\u0026quot; The technological cutting edge was confirmed to be too capital-intensive for most of the industry to adopt immediately. As reported by Narayanage Jayantha, et al. [12], \u0026quot;Hybrid stacking approaches enable incremental AI adoption by integrating with existing rule engines rather than full system replacement.\u0026quot; Incremental integration was validated as a practical, lower-cost adoption pathway.\u003c/p\u003e\n\u003cp\u003eThis theme is highly relevant to EMR\u0026apos;s Technology, Innovation Management, and Entrepreneurship (TIME) department, confirming that cost-effective deployment requires hybridization as an architectural strategy. The novelty is that the most practical adoption pathway involves designing models for interoperability with legacy systems. This is a critical managerial counter-point to the \u0026quot;Big Bang\u0026quot; modernization narratives often associated with new technology. Drawing on the work of Lavika, et al. [1], a full digital transformation requires significant time and investment. The synthesized evidence, however, provided a pragmatic middle ground: using hybrid models to deliver high-impact AI capabilities while mitigating the risk and cost of a total system overhaul [7]. The managerial implication is the prioritization of microservices and API-driven architectures that allow for new AI components to be phased in, thereby avoiding the estimated \u003cimg width=\"27\" height=\"15\" src=\"data:image/png;base64,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\" v:shapes=\"_x0000_i1025\" alt=\"image\"\u003e\u0026nbsp;infrastructure gap and enabling faster time-to-market.\u003c/p\u003e\n\u003ch3\u003eb)\u0026nbsp; \u0026nbsp;\u0026nbsp;Talent and Process Gaps\u003c/h3\u003e\n\u003cp\u003eThe review confirmed that human factors were a major operational barrier (RQ1) due to a misalignment between technical and investigative expertise (Table 5, Sub-Theme 4.2). Consistent with earlier findings by Jovanovic, et al. [48], \u0026quot;Tuning ML models via firefly algorithms reduces hyperparameter optimization time from weeks to hours, addressing skill gaps in model maintenance teams.\u0026quot; Automation was confirmed to mitigate the severity of the talent shortage. Research conducted by Wysocki, et al. [49] provides evidence that \u0026quot;The communication gap between data scientists and fraud investigators causes 40% of high-precision models to be rejected due to unactionable alerts.\u0026quot; Technical accuracy was found not to translate to operational utility without effective cross-functional communication.\u003c/p\u003e\n\u003cp\u003eThis finding is a direct contribution to EMR\u0026apos;s People and Organizations department. The novelty is identifying the Data Scientist-Investigator Communication Gap as a quantifiable cause of operational model rejection. The solution is inherently managerial, not technical. Drawing on the work of Kiran Jot, et al. [62], while comparative evaluations focused on algorithm performance, the synthesized evidence here showed that model value is ultimately determined by \u003cem\u003eorganizational acceptance\u003c/em\u003e. This requires a strategic change management program. Furthermore, in line with the theoretical perspective proposed by Labkoff, et al. [46], a responsible AI framework must include a mandatory step where investigator feedback is formally incorporated into the MLOps training loop. The managerial implication is the necessity of creating a\u0026nbsp;cross-functional MLOps team\u0026nbsp;that includes fraud analysts, ensuring that models are tuned for\u0026nbsp;Actionability\u0026nbsp;and not merely for high precision.\u003c/p\u003e\n\u003ch2\u003eE.\u0026nbsp; \u0026nbsp;\u0026nbsp;Evolving Fraud Ecosystems\u003c/h2\u003e\n\u003cp\u003eThe final theme focused on the dynamic, adversarial nature of financial crime (Table 5, Theme 5), establishing the need for adaptive and robust AI systems that anticipate, rather than merely react to, new threats, which addresses the long-term strategic resilience.\u003c/p\u003e\n\u003ch3\u003ea)\u0026nbsp; \u0026nbsp;\u0026nbsp;Adaptive Adversarial Tactics\u003c/h3\u003e\n\u003cp\u003eThe analysis showed that fraudsters were employing AI to evade detection, necessitating a new defense paradigm (Table 5, Sub-Theme 5.1). Evidence from Wang, et al. [50] suggests that \u0026quot;Generative AI enables synthetic identity fraud at scale, requiring detection systems to shift from pattern recognition to behavioral anomaly detection.\u0026quot; The defense system must adapt to AI-generated, subtle attacks. Consistent with earlier findings by Rocha-Salazar, et al. [23], \u0026quot;Money laundering networks now employ temporal obfuscation techniques that fragment transactions across 15+ accounts within 90-second windows.\u0026quot; This required high-speed, dynamic graph processing for detection.\u003c/p\u003e\n\u003cp\u003eThis theme confirms that the fraud landscape is an AI-driven arms race, directly impacting the strategic planning of the Technology, Innovation Management, and Entrepreneurship (TIME) department. The novelty is the strategic requirement for Adversarial Robustness Testing as a continuous process (RQ1). This goes beyond the periodic model retraining typical in static models. Drawing on the work of Almarshad, et al. [63], the defensive use of Generative Adversarial Networks (GANs) to create synthetic fraud data for model training is demonstrated as a necessary strategic tactic. This contrasts with earlier anomaly detection strategies that assumed a static set of fraudulent behaviors. The managerial implication is the strategic allocation of compute and personnel to pro-actively simulate adversarial attacks, ensuring the long-term operational resilience (RQ1) of the detection system.\u003c/p\u003e\n\u003ch3\u003eb)\u0026nbsp; \u0026nbsp;\u0026nbsp;Regulatory Arbitrage Exploitation\u003c/h3\u003e\n\u003cp\u003eThe review confirmed that fraud rings were exploiting the fragmentation of global regulations (Table 5, Sub-Theme 5.2). The findings of Pocher, et al. [51] indicate that \u0026quot;Decentralized finance platforms create regulatory blind spots where \u003cimg width=\"17\" height=\"15\" src=\"data:image/png;base64,R0lGODlhGgAXAHcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAAABgAZAA0AhAAAAAAAAAAAZgA6kABmtjoAADoAOjo6ZjqQ22YAAGY6AGa222a2/5A6AJBmOrZmALaQOrbb/7b//9uQOtu2Ztv///+2Zv/bkP/btv//tv//2wECAwECAwECAwECAwECAwVuICCKWjNUYwpgThEER6QCU3DOgBUQKBUIkpRFoLjNhkHRI8AYZRILEwqnWjaVhJKRSmoARUjtlEtjip5NMTlnxnZPk+vsYpDrXni8PHUp7FVqOGhkgYANCCkQiCklXwBPX3d5bWcJL1+QQUuTTCEAOw==\" v:shapes=\"_x0000_i1025\" alt=\"image\"\u003e\u0026nbsp;trillion in annual transactions lack standardized monitoring requirements.\u0026quot; The absence of a unified regulatory framework created an exploitable gap. As reported by Kuziemski and Misuraca [52], \u0026quot;Cross-border collaboration mechanisms remain hampered by conflicting data localization laws even among allied nations.\u0026quot; Legal and policy conflicts severely limited global intelligence sharing.\u003c/p\u003e\n\u003cp\u003eThis final theme highlights a key structural challenge for EMR\u0026apos;s Sustainability department: legal friction is the primary bottleneck in global fraud detection, not technical capability. The novelty is identifying Regulatory Arbitrage as a core, persistent threat that technology alone cannot solve. Drawing on the work of Zhang and Zhang [64], a centralized ethics and governance framework is necessary to guide organizations operating in multi-jurisdictional environments. This contrasts with the technical focus of solutions like FL, which work \u003cem\u003earound\u003c/em\u003e the problem by protecting data sovereignty. The managerial implication is the need for senior leadership to drive regulatory advocacy and adopt international standards to harmonize monitoring requirements (RQ3). The strategic decision is to leverage the technological capability of FL (RQ2) while simultaneously advocating for policy changes to close the estimated \u003cimg width=\"24\" height=\"15\" src=\"data:image/png;base64,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\" v:shapes=\"_x0000_i1025\" alt=\"image\"\u003e\u0026nbsp;trillion regulatory gaps.\u003c/p\u003e\n\u003ch2\u003eF.\u0026nbsp; \u0026nbsp;\u0026nbsp;Limitations \u0026amp; Strengths of the Study\u003c/h2\u003e\n\u003cp\u003eA primary limitation of this systematic review was the reliance on publicly accessible scientific literature, potentially excluding proprietary or highly sensitive implementation details and statistics from internal financial institutions\u0026apos; reports. The time and resource constraints limited the inclusion of non-English language publications, which may have led to a geographical bias in the final thematic synthesis. Furthermore, the analysis\u0026apos;s focus on managerial and architectural themes over deep algorithmic performance detail may limit its utility for a purely theoretical computer science audience.\u003c/p\u003e\n\u003cp\u003eThe strengths of the study lie in its direct alignment with the\u0026nbsp;Engineering Management. The use of a\u0026nbsp;Hybrid Inductive-Deductive Thematic Analysis\u0026nbsp;created a novel\u0026nbsp;Managerial Adoption Pathway\u0026nbsp;directly addressing the strategic\u0026nbsp;how-to\u0026nbsp;of AI deployment, a significant gap in existing literature. The synthesis provided a consolidated, evidence-based solution to the long-standing managerial challenges of cross-institutional data collaboration and regulatory explainability (RQ2 and RQ3). The structured PICO/SPIDER frameworks ensured the rigor and reproducibility of the data collection process.\u003c/p\u003e\n\u003ch2\u003eG.\u0026nbsp; \u0026nbsp;Future Research Directions\u003c/h2\u003e\n\u003cp\u003eFuture research should focus on three critical areas to advance the managerial practice of AI deployment. Firstly, an empirical Cost-Benefit Analysis is needed to quantify the Return on Investment (ROI) for Federated Learning architectures versus centralized models, providing financial justification for the strategic CapEx (RQ2). Secondly, future work must develop and validate formal MLOps Maturity Models to assess an organization\u0026apos;s readiness to manage continuous bias monitoring and adaptive adversarial tactics (RQ3). This model would provide a quantifiable benchmark for the organizational factor challenge. Thirdly, experimental studies should investigate the efficacy of Quantum Graph Neural Networks (QGNNs) on standardized financial fraud datasets once the required specialized hardware becomes more commercially viable. This research will prepare engineering managers for the next wave of disruptive computational capability (RQ1).\u003c/p\u003e"},{"header":"IV. CONCLUSION","content":"\u003cp\u003eThe systematic literature review successfully provided a structured, evidence-based roadmap for engineering managers on the strategic adoption of AI in financial fraud detection. The analysis confirmed that operational efficacy (RQ1) required a shift from single-transaction analysis to Network-Aware Detection Architectures like Graph Neural Networks to counter organized crime. Architectural strategy (RQ2) was found to hinge on Distributed Intelligence for Scalable Defense, primarily through Federated Learning to enable privacy-preserving collaboration and Edge-Cloud hybrids to achieve sub-100ms detection latency. Most critically, success depended on Trustworthy AI Governance (RQ3), which mandates integrating Explainable AI for regulatory auditability and continuous bias monitoring. The findings established that the primary non-technical barriers were the high cost of legacy system modernization and the persistent talent gap between data science and fraud investigation. The overall conclusion is that AI deployment in finance is no longer a technical problem but a complex engineering management decision balancing operational speed, data governance, and regulatory compliance.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAcknowledgements:\u003c/h2\u003e\n\u003cp\u003eWe are grateful to all the authors contributing to this research, and we would like to acknowledge the use of Grammarly for language enhancement and correct for this research. No content was taken from AI for this research\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eJ. Lavika, G. Sachin, and D. Parul, \u0026quot;Banks to Lead Digital Transformation With Artificial Intelligence,\u0026quot; in Impact of Artificial Intelligence on Organizational Transformation: Wiley, 2022, pp. 361\u0026ndash;385.\u003c/li\u003e\n\u003cli\u003eH. Rashed, \u0026quot;AI in Financial Services,\u0026quot; in Enterprise Artificial Intelligence Transformation: Wiley, 2020, pp. 67\u0026ndash;84.\u003c/li\u003e\n\u003cli\u003eJ. Jin and Y. Zhang, \u0026quot;The analysis of fraud detection in financial market under machine learning,\u0026quot; Scientific Reports, vol. 15, 2025\u0026ndash;08\u0026ndash;15 2025, doi: 10.1038/s41598-025-15783-2.\u003c/li\u003e\n\u003cli\u003eM. Jullum, A. L\u0026oslash;land, R. B. Huseby, G. \u0026Aring;nonsen, and J. Lorentzen, \u0026quot;Detecting money laundering transactions with machine learning,\u0026quot; Journal of Money Laundering Control, vol. 23, no. 1, pp. 173\u0026ndash;186, 2020\u0026ndash;01\u0026ndash;04 2020, doi: 10.1108/jmlc-07-2019-0055.\u003c/li\u003e\n\u003cli\u003eM. Adil, Z. Yinjun, M. M. Jamjoom, and Z. Ullah, \u0026quot;OptDevNet: A Optimized Deep Event-Based Network Framework for Credit Card Fraud Detection,\u0026quot; IEEE Access, vol. 12, pp. 132421\u0026ndash;132433, 2024, doi: 10.1109/ACCESS.2024.3458944.\u003c/li\u003e\n\u003cli\u003eN. Baisholan, J. E. Dietz, S. Gnatyuk, M. Turdalyuly, E. T. Matson, and K. Baisholanova, \u0026quot;FraudX SimS: A Synthetic Dataset for Anomaly Detection in Payment-Card Transactions,\u0026quot; IEEE Access, vol. 13, pp. 208549\u0026ndash;208562, 2025, doi: 10.1109/ACCESS.2025.3637828.\u003c/li\u003e\n\u003cli\u003eS. K. Hashemi, S. L. Mirtaheri, and S. Greco, \u0026quot;Fraud Detection in Banking Data by Machine Learning Techniques,\u0026quot; IEEE Access, vol. 11, pp. 3034\u0026ndash;3043, 2023, doi: 10.1109/ACCESS.2022.3232287.\u003c/li\u003e\n\u003cli\u003eA. A. Taha and S. J. Malebary, \u0026quot;An Intelligent Approach to Credit Card Fraud Detection Using an Optimized Light Gradient Boosting Machine,\u0026quot; IEEE Access, vol. 8, pp. 25579\u0026ndash;25587, 2020, doi: 10.1109/access.2020.2971354.\u003c/li\u003e\n\u003cli\u003eE. Ileberi, Y. Sun, and Z. Wang, \u0026quot;Performance Evaluation of Machine Learning Methods for Credit Card Fraud Detection Using SMOTE and AdaBoost,\u0026quot; IEEE Access, vol. 9, pp. 165286\u0026ndash;165294, 2021, doi: 10.1109/access.2021.3134330.\u003c/li\u003e\n\u003cli\u003eE. Ileberi, Y. Sun, and Z. Wang, \u0026quot;A machine learning based credit card fraud detection using the GA algorithm for feature selection,\u0026quot; Journal of Big Data, vol. 9, 2022\u0026ndash;02\u0026ndash;25 2022, doi: 10.1186/s40537-022-00573-8.\u003c/li\u003e\n\u003cli\u003eE. Btoush, X. Zhou, R. Gururajan, K. Chan, and O. Alsodi, \u0026quot;Enhancing credit card fraud detection with a stacking-based hybrid machine learning approach,\u0026quot; PeerJ Computer Science, vol. 11, 2025\u0026ndash;09\u0026ndash;02 2025, doi: 10.7717/peerj-cs.3007.\u003c/li\u003e\n\u003cli\u003eD. Narayanage Jayantha, G. D. Dunusinghe, and C. Mrinalini, \u0026quot;Chapter 13 Leveraging Artificial Intelligence for Enhanced Risk Management in Banking: A Systematic Literature Review,\u0026quot; in Artificial Intelligence Enabled Management: An Emerging Economy Perspective: De Gruyter, 2024, pp. 197\u0026ndash;214.\u003c/li\u003e\n\u003cli\u003eR. Roshni, S. Priya, and S. S. Aabha, \u0026quot;The Significance of Generative AI in Enhancing Fraud Detection and Prevention Within the Banking Industry,\u0026quot; in Generative Artificial Intelligence in Finance: Large Language Models, Interfaces, and Industry Use Cases to Transform Accounting and Finance Processes: Wiley, 2025, pp. 159\u0026ndash;173.\u003c/li\u003e\n\u003cli\u003eJ. Jemai, A. Zarrad, and A. Daud, \u0026quot;Identifying Fraudulent Credit Card Transactions Using Ensemble Learning,\u0026quot; IEEE Access, vol. 12, pp. 54893\u0026ndash;54900, 2024, doi: 10.1109/ACCESS.2024.3380823.\u003c/li\u003e\n\u003cli\u003eB. Lebichot, T. Verhelst, Y. A. L. Borgne, L. He-Guelton, F. Obl\u0026eacute;, and G. Bontempi, \u0026quot;Transfer Learning Strategies for Credit Card Fraud Detection,\u0026quot; IEEE Access, vol. 9, pp. 114754\u0026ndash;114766, 2021, doi: 10.1109/ACCESS.2021.3104472.\u003c/li\u003e\n\u003cli\u003eP. H\u0026aacute;jek, M. Z. Abedin, and U. Sivarajah, \u0026quot;Fraud Detection in Mobile Payment Systems using an XGBoost-based Framework,\u0026quot; Information Systems Frontiers, pp. 1\u0026ndash;19, 2022\u0026ndash;10\u0026ndash;14 2022, doi: 10.1007/s10796-022-10346-6.\u003c/li\u003e\n\u003cli\u003eD. Cheng, Y. Zou, S. Xiang, and C. Jiang, \u0026quot;Graph neural networks for financial fraud detection: a review,\u0026quot; Frontiers of Computer Science, vol. 19, 2024\u0026ndash;11\u0026ndash;01 2024, doi: 10.1007/s11704-024-40474-y.\u003c/li\u003e\n\u003cli\u003eS. Luo, X. Chen, Q. Wu, Z. Zhou, and S. Yu, \u0026quot;HFEL: Joint Edge Association and Resource Allocation for Cost-Efficient Hierarchical Federated Edge Learning,\u0026quot; IEEE Transactions on Wireless Communications, vol. 19, no. 10, pp. 6535\u0026ndash;6548, 2020\u0026ndash;02\u0026ndash;26 2020, doi: 10.1109/twc.2020.3003744.\u003c/li\u003e\n\u003cli\u003e\u0026quot;12 XAI-driven Approaches to Detect Financial Frauds and Mitigate Risk,\u0026quot; in AI-driven Cyber Risk Management: River Publishers, 2025, pp. 301\u0026ndash;362.\u003c/li\u003e\n\u003cli\u003eD. Labanca, L. Primerano, M. Markland-Montgomery, M. Polino, M. Carminati, and S. Zanero, \u0026quot;Amaretto: An Active Learning Framework for Money Laundering Detection,\u0026quot; IEEE Access, vol. 10, pp. 41720\u0026ndash;41739, 2022, doi: 10.1109/ACCESS.2022.3167699.\u003c/li\u003e\n\u003cli\u003eB. Stojanović et al., \u0026quot;Follow the Trail: Machine Learning for Fraud Detection in Fintech Applications,\u0026quot; Sensors (Basel, Switzerland), vol. 21, 2021\u0026ndash;02\u0026ndash;25 2021, doi: 10.3390/s21051594.\u003c/li\u003e\n\u003cli\u003eC. Wang, C. Wang, H. Zhu, and J. Cui, \u0026quot;LAW: Learning Automatic Windows for Online Payment Fraud Detection,\u0026quot; IEEE Transactions on Dependable and Secure Computing, vol. 18, no. 5, pp. 2122\u0026ndash;2135, 2021, doi: 10.1109/TDSC.2020.3037784.\u003c/li\u003e\n\u003cli\u003eJ.-D.-J. Rocha-Salazar, M. Segovia‐Vargas, and M. Camacho‐Mi\u0026ntilde;ano, \u0026quot;Money laundering and terrorism financing detection using neural networks and an abnormality indicator,\u0026quot; Expert Syst. Appl., vol. 169, p. 114470, 2021\u0026ndash;05\u0026ndash;01 2021, doi: 10.1016/j.eswa.2020.114470.\u003c/li\u003e\n\u003cli\u003eB. P. Kumar and T. P. Anithaashri, \u0026quot;Novel Patient Monitoring System using Artificial Neural Networks technique comparing with Time Series Analysis,\u0026quot; in 2022 International Conference on Business Analytics for Technology and Security (ICBATS), 16\u0026ndash;17 Feb. 2022 2022, pp. 1\u0026ndash;6, doi: 10.1109/ICBATS54253.2022.9759071. \u003c/li\u003e\n\u003cli\u003eW. Y. B. Lim et al., \u0026quot;Federated Learning in Mobile Edge Networks: A Comprehensive Survey,\u0026quot; IEEE Communications Surveys \u0026amp; Tutorials, vol. 22, no. 3, pp. 2031\u0026ndash;2063, 2019\u0026ndash;09\u0026ndash;26 2020, doi: 10.1109/comst.2020.2986024.\u003c/li\u003e\n\u003cli\u003eJ. Zhang, X. Liu, and P. Zhu, \u0026quot;Cooperation Stimulation and Security in Wireless Ad Hoc Networks \u0026mdash; A Power-Efficient Bayesian Game Approach,\u0026quot; in 2016 IEEE 24th International Symposium on Modeling, Analysis and Simulation of Computer and Telecommunication Systems (MASCOTS), 19\u0026ndash;21 Sept. 2016 2016, pp. 469\u0026ndash;474, doi: 10.1109/MASCOTS.2016.8. \u003c/li\u003e\n\u003cli\u003eA. R. Khalid, N. Owoh, O. Uthmani, M. Ashawa, J. Osamor, and J. Adejoh, \u0026quot;Enhancing Credit Card Fraud Detection: An Ensemble Machine Learning Approach,\u0026quot; Big Data Cogn. Comput., vol. 8, p. 6, 2024\u0026ndash;01\u0026ndash;03 2024, doi: 10.3390/bdcc8010006.\u003c/li\u003e\n\u003cli\u003eM. A. Salam, K. Fouad, D. El-Bably, and S. Elsayed, \u0026quot;Federated learning model for credit card fraud detection with data balancing techniques,\u0026quot; Neural Computing and Applications, vol. 36, pp. 6231\u0026ndash;6256, 2024\u0026ndash;01\u0026ndash;20 2024, doi: 10.1007/s00521-023-09410-2.\u003c/li\u003e\n\u003cli\u003eS. Fritz-Morgenthal, B. Hein, and J. Papenbrock, \u0026quot;Financial Risk Management and Explainable, Trustworthy, Responsible AI,\u0026quot; Front Artif Intell, vol. 5, p. 779799, 2021\u0026ndash;06\u0026ndash;25 2022, doi: 10.3389/frai.2022.779799.\u003c/li\u003e\n\u003cli\u003eJ. M\u0026ouml;kander, M. Axente, F. Casolari, and L. Floridi, \u0026quot;Conformity Assessments and Post-market Monitoring: A Guide to the Role of Auditing in the Proposed European AI Regulation,\u0026quot; Minds and Machines, vol. 32, pp. 241\u0026ndash;268, 2021\u0026ndash;11\u0026ndash;05 2021, doi: 10.1007/s11023-021-09577-4.\u003c/li\u003e\n\u003cli\u003eN. J. Sarna et al., \u0026quot;AI Driven Fraud Detection Models in Financial Networks: A Comprehensive Systematic Review,\u0026quot; IEEE Access, vol. 13, pp. 141204\u0026ndash;141233, 2025, doi: 10.1109/ACCESS.2025.3596060.\u003c/li\u003e\n\u003cli\u003eY. Cheng, Y. Cao, D. Wang, W. Shen, and W. Li, \u0026quot;Boosting Global-Local Feature Matching via Anomaly Synthesis for Multi-Class Point Cloud Anomaly Detection,\u0026quot; IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 12560\u0026ndash;12571, 2025, doi: 10.1109/TASE.2025.3544462.\u003c/li\u003e\n\u003cli\u003eS. Wang, P. Wang, B. Wu, Y. Zhu, W. Luo, and Y. Pan, \u0026quot;Structural entropy minimization combining graph representation for money laundering identification,\u0026quot; International Journal of Machine Learning and Cybernetics, vol. 15, pp. 3951\u0026ndash;3968, 2024\u0026ndash;04\u0026ndash;10 2024, doi: 10.1007/s13042-024-02129-z.\u003c/li\u003e\n\u003cli\u003eF. Wan and P. Li, \u0026quot;A Novel Money Laundering Prediction Model Based on a Dynamic Graph Convolutional Neural Network and Long Short-Term Memory,\u0026quot; Symmetry, vol. 16, p. 378, 2024\u0026ndash;03\u0026ndash;21 2024, doi: 10.3390/sym16030378.\u003c/li\u003e\n\u003cli\u003eS. Ouyang, Q. Bai, H. Feng, and B. Hu, \u0026quot;Bitcoin Money Laundering Detection via Subgraph Contrastive Learning,\u0026quot; Entropy, vol. 26, 2024\u0026ndash;02\u0026ndash;28 2024, doi: 10.3390/e26030211.\u003c/li\u003e\n\u003cli\u003eB. Li et al., \u0026quot;Trustworthy AI: From Principles to Practices,\u0026quot; ACM Computing Surveys, vol. 55, pp. 1\u0026ndash;46, 2021\u0026ndash;10\u0026ndash;04 2021, doi: 10.1145/3555803.\u003c/li\u003e\n\u003cli\u003eG. Peng et al., \u0026quot;BGFL: a blockchain-enabled group federated learning at wireless industrial edges,\u0026quot; Journal of Cloud Computing, vol. 13, 2024\u0026ndash;10\u0026ndash;08 2024, doi: 10.1186/s13677-024-00700-1.\u003c/li\u003e\n\u003cli\u003eJ. Wu, F. Dong, H. Leung, Z. Zhu, J. Zhou, and S. Drew, \u0026quot;Topology-aware Federated Learning in Edge Computing: A Comprehensive Survey,\u0026quot; ACM Computing Surveys, vol. 56, pp. 1\u0026ndash;41, 2023\u0026ndash;02\u0026ndash;06 2023, doi: 10.1145/3659205.\u003c/li\u003e\n\u003cli\u003eX. Jaureguibeitia et al., \u0026quot;Impedance Based Automatic Detection of Ventilations During Mechanical Cardiopulmonary Resuscitation,\u0026quot; in 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 23\u0026ndash;27 July 2019 2019, pp. 19\u0026ndash;23, doi: 10.1109/EMBC.2019.8856822. \u003c/li\u003e\n\u003cli\u003eL. Li, L. Zhu, and W. Li, \u0026quot;Cloud\u0026ndash;Edge\u0026ndash;End Collaborative Federated Learning: Enhancing Model Accuracy and Privacy in Non-IID Environments,\u0026quot; Sensors (Basel, Switzerland), vol. 24, 2024\u0026ndash;12\u0026ndash;01 2024, doi: 10.3390/s24248028.\u003c/li\u003e\n\u003cli\u003eJ. Zhang, R. Ding, M. Ban, and L. Dai, \u0026quot;PKU-GoodsAD: A Supermarket Goods Dataset for Unsupervised Anomaly Detection and Segmentation,\u0026quot; IEEE Robotics and Automation Letters, vol. 9, no. 3, pp. 2008\u0026ndash;2015, 2024, doi: 10.1109/LRA.2024.3352358.\u003c/li\u003e\n\u003cli\u003eA. Kuznietsov, B. Gyevnar, C. Wang, S. Peters, and S. Albrecht, \u0026quot;Explainable AI for Safe and Trustworthy Autonomous Driving: A Systematic Review,\u0026quot; IEEE Transactions on Intelligent Transportation Systems, vol. 25, pp. 19342\u0026ndash;19364, 2024\u0026ndash;02\u0026ndash;08 2024, doi: 10.1109/tits.2024.3474469.\u003c/li\u003e\n\u003cli\u003eJ. Mu\u0026ntilde;oz-Ord\u0026oacute;\u0026ntilde;ez, C. Cobos, J.-C. Vidal-Rojas, and F. Herrera, \u0026quot;A Maturity Model for Practical Explainability in Artificial Intelligence-Based Applications: Integrating Analysis and Evaluation (MM4XAI-AE) Models,\u0026quot; Int. J. Intell. Syst., vol. 2025, 2025\u0026ndash;01\u0026ndash;01 2025, doi: 10.1155/int/4934696.\u003c/li\u003e\n\u003cli\u003eN. Balasubramaniam, M. Kauppinen, A. Rannisto, K. Hiekkanen, and S. Kujala, \u0026quot;Transparency and explainability of AI systems: From ethical guidelines to requirements,\u0026quot; Inf. Softw. Technol., vol. 159, p. 107197, 2023\u0026ndash;03\u0026ndash;01 2023, doi: 10.1016/j.infsof.2023.107197.\u003c/li\u003e\n\u003cli\u003eE. Ferrara, \u0026quot;Should ChatGPT be Biased? Challenges and Risks of Bias in Large Language Models,\u0026quot; First Monday, vol. 28, 2023\u0026ndash;04\u0026ndash;07 2023, doi: 10.5210/fm.v28i11.13346.\u003c/li\u003e\n\u003cli\u003eS. Labkoff et al., \u0026quot;Toward a responsible future: recommendations for AI-enabled clinical decision support,\u0026quot; Journal of the American Medical Informatics Association : JAMIA, vol. 31, pp. 2730\u0026ndash;2739, 2024\u0026ndash;09\u0026ndash;26 2024, doi: 10.1093/jamia/ocae209.\u003c/li\u003e\n\u003cli\u003eN. Innan et al., \u0026quot;Financial fraud detection using quantum graph neural networks,\u0026quot; Quantum Machine Intelligence, vol. 6, 2023\u0026ndash;09\u0026ndash;03 2023, doi: 10.1007/s42484-024-00143-6.\u003c/li\u003e\n\u003cli\u003eD. Jovanovic, M. Antonijevic, M. Stanković, M. Zivkovic, M. Tanaskovic, and N. Bačanin, \u0026quot;Tuning Machine Learning Models Using a Group Search Firefly Algorithm for Credit Card Fraud Detection,\u0026quot; Mathematics, 2022\u0026ndash;06\u0026ndash;29 2022, doi: 10.3390/math10132272.\u003c/li\u003e\n\u003cli\u003eO. Wysocki et al., \u0026quot;Assessing the communication gap between AI models and healthcare professionals: explainability, utility and trust in AI-driven clinical decision-making,\u0026quot; Artif. Intell., vol. 316, p. 103839, 2022\u0026ndash;04\u0026ndash;11 2022, doi: 10.1016/j.artint.2022.103839.\u003c/li\u003e\n\u003cli\u003eY. C. Wang, J. Xue, C. Wei, and C. c. J. Kuo, \u0026quot;An Overview on Generative AI at Scale With Edge\u0026ndash;Cloud Computing,\u0026quot; IEEE Open Journal of the Communications Society, vol. 4, pp. 2952\u0026ndash;2971, 2023\u0026ndash;06\u0026ndash;02 2023, doi: 10.1109/ojcoms.2023.3320646.\u003c/li\u003e\n\u003cli\u003eN. Pocher, M. Zichichi, F. Merizzi, M. Shafiq, and S. Ferretti, \u0026quot;Detecting anomalous cryptocurrency transactions: An AML/CFT application of machine learning-based forensics,\u0026quot; Electronic Markets, vol. 33, pp. 1\u0026ndash;17, 2022\u0026ndash;06\u0026ndash;07 2022, doi: 10.1007/s12525-023-00654-3.\u003c/li\u003e\n\u003cli\u003eM. Kuziemski and G. Misuraca, \u0026quot;AI governance in the public sector: Three tales from the frontiers of automated decision-making in democratic settings,\u0026quot; Telecomm Policy, vol. 44, no. 6, p. 101976, Jul 2020, doi: 10.1016/j.telpol.2020.101976.\u003c/li\u003e\n\u003cli\u003eS. Xiang, G. Zhang, D. Cheng, and Y. Zhang, \u0026quot;Enhancing Attribute-Driven Fraud Detection With Risk-Aware Graph Representation,\u0026quot; IEEE Transactions on Knowledge and Data Engineering, vol. 37, no. 5, pp. 2501\u0026ndash;2512, 2025, doi: 10.1109/TKDE.2025.3543887.\u003c/li\u003e\n\u003cli\u003eG. Duan, H. Lv, H. Wang, G. Feng, and X. Li, \u0026quot;Practical Cyber Attack Detection With Continuous Temporal Graph in Dynamic Network System,\u0026quot; IEEE Transactions on Information Forensics and Security, vol. 19, pp. 4851\u0026ndash;4864, 2024, doi: 10.1109/TIFS.2024.3385321.\u003c/li\u003e\n\u003cli\u003eC. Lawal, O. O. Moses, K. Okokpujie, E. Adetiba, B. Adebisi, and D. O. Emmanuel, \u0026quot;FedHiGraphSAGE: A Hierarchical Graph Neural Network Model for Multi-Level Anomaly Detection in Federated Cloud Environments,\u0026quot; IEEE Access, vol. 13, pp. 183985\u0026ndash;184010, 2025, doi: 10.1109/ACCESS.2025.3622335.\u003c/li\u003e\n\u003cli\u003eM. Li, L. Jia, and X. Su, \u0026quot;Global-local graph attention with cyclic pseudo-labels for bitcoin anti-money laundering detection,\u0026quot; Scientific Reports, vol. 15, 2025\u0026ndash;07\u0026ndash;02 2025, doi: 10.1038/s41598-025-08365-9.\u003c/li\u003e\n\u003cli\u003eT. Ashfaq et al., \u0026quot;A Machine Learning and Blockchain Based Efficient Fraud Detection Mechanism,\u0026quot; Sensors (Basel, Switzerland), vol. 22, 2022\u0026ndash;09\u0026ndash;21 2022, doi: 10.3390/s22197162.\u003c/li\u003e\n\u003cli\u003eD. Nguyen et al., \u0026quot;Federated Learning Meets Blockchain in Edge Computing: Opportunities and Challenges,\u0026quot; IEEE Internet of Things Journal, vol. 8, pp. 12806\u0026ndash;12825, 2021\u0026ndash;04\u0026ndash;05 2021, doi: 10.1109/jiot.2021.3072611.\u003c/li\u003e\n\u003cli\u003eH. G. Abreha, M. Hayajneh, and M. Serhani, \u0026quot;Federated Learning in Edge Computing: A Systematic Survey,\u0026quot; Sensors (Basel, Switzerland), vol. 22, 2022\u0026ndash;01\u0026ndash;01 2022, doi: 10.3390/s22020450.\u003c/li\u003e\n\u003cli\u003eJ. Černevičienė and A. Kaba\u0026scaron;inskas, \u0026quot;Explainable artificial intelligence (XAI) in finance: a systematic literature review,\u0026quot; Artificial Intelligence Review, vol. 57, 2024\u0026ndash;07\u0026ndash;26 2024, doi: 10.1007/s10462-024-10854-8.\u003c/li\u003e\n\u003cli\u003eN. Corr\u0026ecirc;a et al., \u0026quot;Worldwide AI ethics: A review of 200 guidelines and recommendations for AI governance,\u0026quot; Patterns, vol. 4, 2022\u0026ndash;06\u0026ndash;23 2022, doi: 10.1016/j.patter.2023.100857.\u003c/li\u003e\n\u003cli\u003eS. Kiran Jot et al., \u0026quot;Comparative Evaluation of Machine Learning Algorithms for Bank Fraud Detection,\u0026quot; in Integrating Metaheuristics in Computer Vision for Real-World Optimization Problems: Wiley, 2024, pp. 71\u0026ndash;79.\u003c/li\u003e\n\u003cli\u003eF. A. Almarshad, G. A. Gashgari, and A. I. A. Alzahrani, \u0026quot;Generative Adversarial Networks-Based Novel Approach for Fraud Detection for the European Cardholders 2013 Dataset,\u0026quot; IEEE Access, vol. 11, pp. 107348\u0026ndash;107368, 2023, doi: 10.1109/ACCESS.2023.3320072.\u003c/li\u003e\n\u003cli\u003eJ. Zhang and Z.-M. Zhang, \u0026quot;Ethics and governance of trustworthy medical artificial intelligence,\u0026quot; BMC Medical Informatics and Decision Making, vol. 23, 2023\u0026ndash;01\u0026ndash;13 2023, doi: 10.1186/s12911-023-02103-9.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"University of Essex","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":"Engineering Management, Artificial Intelligence, Fraud Detection, Federated Learning, Explainable AI (XAI), Organizational Strategy, Governance, Regulatory Compliance, Distributed Computing, Graph Neural Networks","lastPublishedDoi":"10.21203/rs.3.rs-9267907/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9267907/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe financial sector's transition to AI-driven fraud detection requires new strategies for technology adoption that balance operational efficacy, data privacy, and strict regulatory governance. Existing literature often focuses on algorithmic performance, neglecting the critical architectural and managerial implications. The study aimed to systematically review the literature to provide an evidence-based pathway for engineering managers concerning the operationalization, architecture, and governance of AI-powered fraud detection systems (RQ1-RQ3). A Systematic Literature Review (SLR) was conducted across IEEE Xplore, Scopus, and other high-impact databases (2000–2026). A hybrid thematic analysis was performed on the extracted data to develop a Grounded Theory of AI Adoption Pathways. The analysis established that operational efficacy (RQ1) relies on Graph Neural Networks to counter complex fraud rings. Architectural strategy (RQ2) mandates Distributed Intelligence, with Federated Learning confirmed as a tool for collaboration under GDPR constraints and Edge-Cloud hybrids achieving sub-100ms latency. Governance (RQ3) requires mandatory XAI for regulatory compliance and continuous bias monitoring, with bias audits revealing a 22% higher false positive rate in some cross-border transactions. Successful AI deployment is a strategic management decision requiring new architectural and governance frameworks to ensure regulatory trust and operational resilience. Future empirical research should focus on a formal MLOps Maturity Model to benchmark organizational readiness for bias mitigation and adversarial robustness.\u003c/p\u003e","manuscriptTitle":"AI-Powered Fraud Detection in Financial Networks: A Systematic Literature Review.","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-31 06:47:54","doi":"10.21203/rs.3.rs-9267907/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":"21dacf9d-f0cc-4139-9f7b-6ec2f139a9ac","owner":[],"postedDate":"March 31st, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-03-31T06:47:54+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-31 06:47:54","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9267907","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9267907","identity":"rs-9267907","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}