Explainable and Adversarial Robust Deep Learning for Malware Campaigns Forensic Attribution

Explainable and Adversarial Robust Deep Learning for Malware Campaigns Forensic Attribution | 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 Research Article Explainable and Adversarial Robust Deep Learning for Malware Campaigns Forensic Attribution Idowu Olugbenga ADEWUMI, Wumi AJAYI, Tolulope OLUFEMI, Ayoade Oluwafisayo BABATOPE This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8515849/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 This research paper proposes a machine learning framework for malware attribution in digital forensics based on Random Forest. The model takes malware samples as input and classifies them into attribution labels such as APT1, APT28, APT33, CyberGangX. These features include File_Size_KB, Num_Functions, Num_Imports, and Entropy. The model was trained on a dataset consisting of 5000 samples. The important features have a wide range of values. The feature File_Size_KB varies from 11 KB to 4998 KB. Likewise, entropy varies from 1.5 to 8. The model's accuracy was found to be 19.7% with precision, recall and F1 scores at 20% on average overall attribution. According to the feature importance plot of the Random Forest model, the key features were File_Size_KB, Num_Functions, and Entropy. In terms of methodology, the model was trained using the normal method, whereby 80% of the dataset was used for training and 20% for testing. The features were normalized using StandardScaler and labels were converted from categorical to numerical values using LabelEncoder. A decline in performance was observed during adversarial robustness testing, with F1-score dropping from 20% on clean data to 15% on adversarially perturbed data. The model has difficulty due to class imbalance causing it to misclassify classes mostly represented, like CyberGangX and Unknown. The model produced higher results for some classes (e.g. APT1), it had low precision and recall for many other classes. The authors mention another challenge of robustness; the model can be fooled with a small perturbation. To sum up, while the proposed model can strengthen the existing malware attribution processes, its scalability, performance, and adversarial defenses need improvement. To improve robustness in our future efforts, we should focus on hyperparameter tuning, picking better models and trying the adversarial training. Further, feature engineering and network traffic data inclusion can enhance the model performance by increasing the accuracy and by allowing it to classify more complex malware. The findings of the study indicate that it is very important to develop a good predictive model that can be interpreted. Further, it will help cybersecurity professionals and law enforcement agencies in the field of digital forensics. Artificial Intelligence and Machine Learning Malware Attribution Machine Learning Random Forest Adversarial Robustness Digital Forensics Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 1. Introduction Malware attribution is an essential digital forensics tool aimed at identifying and analyzing the malware used in cyberattacks. Malware attribution is the process of linking malware to cybercriminals and their motives. Grier et al. ( 2013 ) observed that this process helps forensic investigators establish the identity of a malware and the damage it intends to cause. Effective malware attribution is important not only for identifi- cation, but also for preventing future attacks, ensuing legal action, and enhancing cybersecurity. Although important, conventional malware attribution methods are limited in scope due to their often use of manual analysis and signature-based detection systems, which are ineffective against sophisticated and evasive malware strains (Zhou et al., 2020 ). Consequently, these methods fail to adapt to the new strategies employed by attackers, especially considering the use of obfuscation and polymorphism by contemporary malware programs. The scalability and efficiency of traditional malware attribution are one of the main issues especially when there is a huge amount of data or complex attacks. Signature-based systems can only detect known attacks. Thus, zero-day attacks and new variations of known malware remain undetected by these types of systems (Xia et al., 2018 ). In addition, they do not provide useful details on how threat actors behave and attack. As cyberattacks continue to evolve, there is a rising demand for machine learning methods that are both explainable and robust that can overcome the limitations of standard ML to get accurate and explainable forensic investigations. Some of the malware attribution models that are present today have big gaps particularly in their interpretability and adversarial robustness. Numerous machine learning models currently in existence are “black boxes,” meaning that decisions made by these models cannot be easily understood by an expert human (Gilpin et al. 2018 ). The absence of transparency creates hurdles in forensic contexts where knowing the reason behind a model’s decision is crucial for acceptance in other processes. These models are also susceptible to adversarial attacks which can interfere with or deceive attribution, thus compromising the reliability of their forensic findings (Goodfellow et al., 2015 ). Thus, the main research problem of the study is to develop a machine learning-based malware attribution model which is explainable and adversarially defensive. This research aims to achieve two objectives. Initially, the purpose of the study was to establish a deep learning model for evidence-based attribution of malware campaigns with explainability. Secondly, we seek to make the attribution model robust against adversarial attacks for a safe application in forensic tests in real life. This research aims to address these objectives in order to provide a bridge between complex models of machine learning and the design requirements of digital forensics. The goal of this study is to produce a suitable explainable and adversarially robust malware attribution system for digital forensics. As machine learning techniques are integrated into digital forensics it is important that systems do not just perform well, but render explanations such that forensic experts, law enforcement and the judiciary can understand them ( Ribeiro et al. 2016 ). The importance of explainability for machine learning models is crucial in a forensic setting, where the evidence admissibility in court depends on whether the explainability for the methods was justified (Lakkaraju et al., 2019 ). This study can greatly affect many other studies. This provides a tool for law enforcement for better targeting of resources and response or attribution of cyberattacks. Cybersecurity professionals may benefit from the framework as it helps in identifying threats as they emerge. In addition, legal stakeholders will be able to use the proposed system’s transparency and adversarially resistance to lend credibility to their court claims, thus backing the reliability of digital evidence. If software attribution methods are made more accurate, transparent, and robust, it will contribute to secure and trustworthy digital forensics. 2. Related Work Attribution of malware is paramount to digital forensics as it helps investigators trace the source of an attack. The existing attribution methodology of malware relies on the conventional signature-based software, which employs specific features of the known attack like similar code signature and hash value (Zhou et al., 2020 ). While these techniques are effective in detecting known malware, they are fundamentally limited in detecting new or polymorphic malware, which adeptly alters its appearance to elude backdated detection Zhou et al., 2019 . One traditional strategy is behavioral analysis, malware behaviour is monitored during the execution phase for suspicious patterns (Cui et al., 2018 ). While they are more flexible, behavioral analysis’s high computational and temporal cost to assess large datasets, makes them unsuitable for real-time forensic investigations. In the last few years, malware attribution systems have been increasingly equipped with machine learning (ML) methods to combat these limitations. Machine learning (ML) methodology, especially from supervised learning, has been found effective in classifying malware based on their features like system calls, byte strings, and network traffic (Xia et al., 2018 ). Even though they have strong points, these methods could not help an individual in some cases due to issues like lack interpretability, attack to adversaries, and so on. As a consequence, Grier et al. 2013 stressed a need for more robust and explainable ML models that provide accurate results and transparent decision making, particularly in forensic contexts where legal admissibility is a concern. Explainability in machine learning (XAI) is essential in critical areas such as digital forensics where automated decisions affect legal proceedings, law enforcement action, and cybersecurity response. In investigation, being able to explain why was a particular choice made, which means forensic must justify findings in court. Many current ML models, particularly the deep learning-based ones, are “black boxes” that give little to no information on how a decision was made (Gilpin et al., 2018 ). The inability to verify the system’s predictions makes it impossible to trust such a system. This would not be good in a legal and investigative situation. Various explainable AI techniques have been designed to address this empirical gap in interpretability of machine learning models. Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) have been used to provide local explanations for individual prediction using which features were most responsible for prediction (Ribeiro et al., 2016 ). In the area of malware detection and attribution, explainability helps demonstrate which features – for example, certain byte patterns or system behaviours – were responsible for the model classifying a sample as malicious. But although these methods help to make images interpretable, they do not solve the challenge of robustness against adversarial manipulation, which is crucial in digital forensics. Cybersecurity and digital forensics heavily rely on machine learning systems, and adversarial attacks specially target these systems, thereby compromising the evidence integrity. Adversarial attacks are the perturbation of the input with tiny changes that result in the model making incorrect predictions. Moreover, the input is changed in a way that is not visible to the human eye. (Goodfellow et al., 2015 ) Adversarial attacks can confuse the model in malicious attribution which can lead to misclassification of benign files as malicious and vice versa. This issue is particularly worrisome when the use of these machine learning models is for forensic application where the legal consequence needs utmost confidence in their reliability. Strategies to apply in defending against adversarial attacks are being proposed. One way is adversarial training, which augments the training dataset with adversarial examples in order to render the model more robust to such perturbations (Madry et al., 2017 ). Methods such as adversarial detection which focus on identifying and denying these inputs before they reach the model (Papernot et al., 2017 ). In malware attribution models within digital forensics, adversarial robustness is must that adversaries cannot manipulate the malware attribution models’ predictions. Additionally, adversarially robust models should be explainable as the forensic expert must not only trust the model’s prediction but must also know how the model has come to that prediction especially when there could be adversarial inputs that change the behaviour of the model. 3. Methodology 3.1 Framework Overview The proposed deep learning framework for malware attribution is characterized by two important properties required for digital forensics explainability and adversarial robustness. Within the field of digital forensics, where accuracy and transparency are fundamental, explainability provides insight on the reasoning and interpretation of models. By putting explainability techniques like LIME and SHAP we can find out what feature led to the classification of the malware sample. The clarity this offers of the decision taken by the model will inform forensic investigators and lawyers (Ribeiro et al., 2016 ). At the same time, the framework also focuses on adversarial robustness which makes them resilient to such attacks. Malware samples can be subtly modified so that their predictions can be manipulated. Adversarial training is key. Using adversarial training methods such as the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) ensures that the model is taught to learn and withstand adversarial perturbations, thereby allowing it to remain accurate in real-world forensic scenarios (Goodfellow et al., 2015 ). 3.2 Model Architecture The core architecture of the deep learning framework combines Convolutional Neural Networks with Recurrent Neural Network and Transformer. This hybrid high-performance model allows for effective malware classification and attribution. The CNNs are used for feature extraction from the malware binary or its byte sequence and identifies different patterns that are important to separate malware families. The CNN layer can capture spatial hierarchies effectively, which can be useful for recognizing system calls and byte sequences that are structured differently across various malware types. We use RNNs, specifically LSTM networks, to model the sequential relationships in the malware execution process. They are especially well-suited for capturing temporal dependencies, like the order in which malware invokes system calls or issues network packets during execution. By using LSTMs, the model is able to learn the sequence of events that led to a malicious behaviour. This is very important for attributing the malware campaigns to specific threat actors based on their behaviour over time. Transformers work on large amount of data and capture long range dependencies which cannot be captured by CNNs and RNNs. The incorporation of attention mechanism in transformers leads to improved accuracy for classifications and also helps interpret the results and features. The malware’s approach execution is made focusable for this very reason. The use of LIME or SHAP improves the model’s explainability by providing local explanation for every individual prediction of the ML model. These techniques generate feature importance scores for the various inputs (e.g., byte sequences, system calls) explaining how each feature helped in the final classification. Besides, the adversarial training occurs through FGSM and PGD so that they can maintain their reliability even when proven with the help of adversarial samples. Table 1 Overview of the Proposed Deep Learning Component Description Purpose Explainability Integration of explainable AI (XAI) techniques such as LIME and SHAP to provide interpretability of model decisions. Ensures transparency by explaining the reasons behind malware classifications, crucial for forensic investigations. Adversarial Robustness Application of adversarial training techniques (e.g., FGSM, PGD) to improve model resilience to adversarial attacks. Enhances the model's ability to withstand manipulation, ensuring the robustness of the forensic system. Machine Learning Techniques Use of deep learning models, including CNNs (for feature extraction), RNNs (for sequential dependencies), and Transformers (for capturing long-range dependencies). Combines powerful models to extract patterns from diverse data types and improve classification accuracy. Malware Attribution Classifies and attributes malware families based on features like system calls, byte sequences, and execution patterns. Provides precise attribution to aid in identifying threat actors and understanding malware campaigns. Forensic Application Designed to provide actionable insights for law enforcement, cybersecurity professionals, and legal stakeholders. Facilitates the use of the model’s predictions in real-world forensic and legal contexts. Table 2 Feature Extraction and Preprocessing Pipeline Step Description Tools Used Purpose Feature Extraction Extraction of raw features from malware samples, including system calls, byte sequences, and API calls. Cuckoo Sandbox, Yara, PE Studio Extract relevant data from malware for input into the model, focusing on behavior and content. Normalization Scaling features to a standard range (e.g., 0–1) to ensure that all features contribute equally to the model. Python (scikit-learn), TensorFlow Normalize extracted features to improve model convergence and prevent dominance of any feature. Transformation Applying transformations to data, such as one-hot encoding for categorical data and embedding for sequences. Python (pandas, Keras), scikit-learn Transform categorical and sequential data into formats suitable for deep learning models. Data Augmentation Generation of synthetic malware samples through adversarial perturbation or data synthesis techniques. Adversarial libraries (e.g., CleverHans) Expand the dataset and improve model robustness by simulating new, unseen data variations. Feature Selection Identifying and selecting the most informative features using methods like Recursive Feature Elimination (RFE). scikit-learn, L1 Regularization Reduce dimensionality by selecting the most relevant features, improving training efficiency. Noise Removal Removal of irrelevant or noisy data points to ensure that the model is trained on clean data. Python (NumPy, pandas), custom scripts Clean the data by eliminating anomalies and irrelevant data that could affect model performance. Table 3 Adversarial Training Techniques Technique Description Role in Enhancing Model Robustness Purpose Fast Gradient Sign Method (FGSM) A method that generates adversarial examples by adjusting input data in the direction of the gradient of the loss with respect to the input. FGSM is quick and efficient, providing a simple way to introduce adversarial perturbations to the model. To enhance robustness by training the model to resist small perturbations and avoid misclassification under adversarial attacks. Projected Gradient Descent (PGD) A more advanced adversarial method that iteratively adjusts input data based on gradient steps, projecting it back into a feasible domain. PGD is stronger than FGSM as it performs iterative optimization, generating more powerful adversarial examples. To improve model resilience against stronger adversarial attacks by forcing the model to learn from more difficult adversarial examples. Carlini & Wagner (C&W) Attack An optimization-based method that generates adversarial samples by solving a constrained optimization problem to minimize perturbation while ensuring misclassification. C&W is one of the most effective attacks, producing minimal perturbations that are hard for the model to detect. To test and enhance the model’s robustness against highly sophisticated adversarial strategies. DeepFool An iterative method that finds the minimal perturbation needed to change the classification of the input. DeepFool focuses on finding the smallest perturbation, providing a powerful test of model vulnerability. To challenge the model’s robustness by applying minimal perturbations and assessing its ability to correctly classify under such conditions. Adversarial Training A method of training the model on both clean and adversarial examples to improve its ability to correctly classify adversarially perturbed inputs. Incorporates adversarial examples into training to improve the model's performance on adversarial inputs, enhancing general robustness. To directly train the model to become more resistant to adversarial attacks by exposing it to adversarial examples during the learning process. Table 4 Evaluation Metrics Metric Description Purpose Accuracy The proportion of correctly classified samples (both benign and malicious) over the total number of samples. Measures the overall performance of the model, indicating how well it classifies both malware and benign files. F1 Score The harmonic mean of precision and recall, providing a balance between them, especially useful for imbalanced datasets. Provides a balanced performance measure when dealing with imbalanced classes, ensuring the model performs well on both positive and negative classes. Precision The proportion of true positives (correctly identified malware samples) out of all positive predictions made by the model. Evaluates how many of the predicted malware samples are truly malicious, minimizing false positives. Recall The proportion of true positives out of all actual positive samples (malicious files). Measures the model’s ability to identify all malicious samples, focusing on minimizing false negatives. Adversarial Robustness The model’s ability to maintain high performance when adversarial perturbations are applied to the input data. Tests the resilience of the model under adversarial conditions, ensuring that it can still classify correctly when manipulated. Interpretability Score A measure of how easily the model's decision-making process can be understood by human experts (using techniques like SHAP or LIME). Assesses the transparency of the model, ensuring that its predictions are explainable and usable in forensic contexts. Area Under the ROC Curve (AUC-ROC) The area under the receiver operating characteristic curve, representing the trade-off between true positive rate and false positive rate. A measure of the model's ability to discriminate between classes, with a higher AUC indicating better overall performance. Confusion Matrix A table showing the true positives, true negatives, false positives, and false negatives of the model’s predictions. Provides a detailed breakdown of the model’s predictions, helping to identify areas of improvement in classification. Robustness to Perturbations Evaluates how well the model performs under different adversarial attacks, such as FGSM, PGD, and others. Quantifies the model's stability and resilience to attacks, ensuring that it maintains accuracy in adversarial conditions. The proposed deep learning framework consists of a number of essential components that guarantee both the robustness and reliability of the malware attribution system. The framework highlight explainability using tools LIME and SHAP to explain the reasons for classifying the model and bring transparency. This is essential in criminal investigations where the reasoning behind predictions needs to be shown in a court of law. Moreover, the framework integrates adversarial robustness, integrating adversarial training techniques such as the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) towards improving the model’s ability to withstand the effect of adversarial attacks, ensuring reliability under adversarial inputs. By utilizing various machine learning techniques, including CNNs, RNNs, and Transformers, the model can learn complex representations from various data sources, allowing for improved malware characterization and attribution performance. Malware attribution involves sorting malware to families, based on system calls, byte sequences, and the actual execution style. This classifications improves understanding of the motivation behind the malware campaign. To sum up, running the framework can be used with a forensic application which can be used practically for law enforcement. The Feature Extraction and Preprocessing Pipeline provides a detailed overview of the critical steps taken to prepare malware for analysis. The first feature extraction is done through tools like Cuckoo Sandbox, Yara, PE Studio where they capture system calls, byte sequences and API calls. The normalisation of the extracted feature is done through the scikit-learn or TensorFlow python libraries so that all features contribute equally to the model. In addition, one-hot coding for categorical data and embedding for sequences are examples of transformations employed for formatting the data to be suitable for deep learning data. The ability to build a model which fine-tunes images using computer vision techniques will go a long way to prevent accident. Data augmentation through adversarial perturbations and feature selection technique like Recursive Feature Elimination (RFE) will also help in making the model robust and efficient through extension in data and focusing on important features. This refers to application of noise removal techniques in order to clean the data. The section on Adversarial Training Techniques compares different approaches for improving model robustness. FGSM, PGD, Carlini & Wagner (C&W) assault as well as DeepFool are all antagonistic methodologies applied to assess the model’s sensitivity to perturbations. These procedures are important so that the model works well in situations out in the world where enemy attacks will be likely to happen. By utilizing both clean and adversarial examples in the training of the model, adversarial training allows the model to resist such attacks, thus enhancing the model’s overall robustness. In the end, evaluation metrics help assess a model’s performance, interpretability, and robustness. Choosing an accurate metric can be taxing sometimes. Some models may perform well in terms of accuracy but may perform poorly in terms of precision. The adversarial robustness score assesses how well the model maintains its performance when working with adversarial examples, while the interpretation score measures its ability to explain its prediction as a function of the input. The AUC-ROC refers to Area Under the Receiver Operating Characteristic Curve. This metric indicates the model’s ability to distinguish between classes. At the same time, the confusion matrix provides a comprehensive account of the model’s predictions. The robustness of the model to perturbations indicates the model’s resilience to adversarial perturbations. Thus, the attacker will not have much effect on the quality of the model’s prediction when subjected to adversarial perturbations. The study methodology leverages advanced deep learning techniques and rigorously chosen preprocessing, adversarial training, and evaluation metrics applicable to forensic usability in malware attribution. 3.3 Data Collection and Preprocessing The research will incorporate actual malware data sets in training and evaluating the suggested deep learning model for this study. Noteworthy datasets like EMBER (Endgame Malware Benchmark for Enhanced Research) and Malimg (Malware Image Dataset) will serve as sources for malware samples, each labelled with its malware family or classification (Grosse et al., 2017). The cybersecurity and digital forensics communities widely utilize these datasets to ensure the model is trained using a variety of malware samples with different characteristics. Thus, increasing the model’s generalization ability. Transformations on the data set to prepare the data for deep learning models are data preprocessing. The malware’s raw file will be converted to structure form data which is in byte code, system calls, API calls, and network traffic. Using several static and dynamic analysis tools, like Cuckoo Sandbox for malware behavioral analysis, features like frequency of system call and frequency of byte will be extracted. The features are normalized so that they are on the same scale. This is performed so that every feature is treated equally during the training of a model. To represent the features effectively, transformation techniques will also be used. For example, categorical variables (file type, malware family, etc.) will be one-hot encoded. The byte sequence will be embedded. The preprocessing pipeline makes sure the input data to the model is clean as well as suitable for deep learning ensuring the accuracy of models, and reducing training time. Table 5 Data Splitting Strategy Step Description Method/Technique Purpose Dataset Division Splitting the dataset into three subsets: training, validation, and test sets. 70% Training, 15% Validation, 15% Test Ensures that the model is trained on a large portion of the data while having separate subsets for model evaluation and hyperparameter tuning. Stratified Sampling Ensuring that the distribution of malware families is maintained in all subsets. Stratified Sampling Prevents class imbalance issues by ensuring that each subset has an equal representation of each malware family. Random Sampling Randomly selecting samples for the training, validation, and test sets, while preserving class distribution. Random Sampling (with stratification) Guarantees that all data points are equally likely to be selected for any subset, promoting fairness and randomness in model training and testing. Cross-Validation Setup Dividing the dataset into k folds for cross-validation to reduce variance in performance estimates. 5-Fold or 10-Fold Cross-Validation Helps assess the model's performance on different portions of the dataset, ensuring reliable results by using each subset as both training and testing data. Data Shuffling Randomly shuffling the dataset before splitting to prevent any ordering bias. Shuffling before splitting Ensures that the model does not learn patterns from the order of the data, improving generalization. 3.4 Evaluation Metrics The effectiveness of the recommended model will be evaluated using a number of measures; both for performance and interpretability. The performance of the model in the malware classification objective is measured mainly using accuracy and F1 score. The importance of F1 score is to allow the precision and recall to be balanced in our case. It is especially helpful in the case of imbalanced datasets where the benign (non-malware) files may be overrepresented in the dataset. To gauge the model’s competence, the Precision and Recall will be evaluated. Precision assesses the ability to classify malicious samples accurately and Recall to cover the entire range of malicious samples. The model will be tested against adversarial attacks such as FGSM and PGD to evaluate its Adversarial Robustness. A comparison of the model’s performance on adversarially perturbed samples compared to clean data will exhibit robustness to adversarial manipulation of model inputs. The metrics basically focus on the use of LIME and SHAP for quantifying the interpretability of model predictions. This will involve measuring how stable the feature importance scores are and how interpretable the explanations are for varying malware variants. A higher interpretability score means forensic experts can confidently use the model for attribution tasks. Using a combination of these metrics, a comprehensive evaluation of both the performance and applicability of the model for digital forensics can take place, thereby accommodating the need for both robustness and transparency. 4. Experimental Setup 4.1 Dataset Description For this paper, the study utilize a combination of malware datasets which are used widely in the digital forensics and cyber security researchers’ community. The datasets further provide different kinds of malware samples that come with labels for identifying the malware families or types. These labels are essential to training as well as evaluating the deep learning model. The study employs two major datasets EMBER and Malimg. Embers is also a malware dataset and specifically geared towards malware classification. Malimg is specifically a malware attribution dataset. The EMBER dataset includes more than one million features extracted from Windows executable files. It contains comprehensive information about the metadata of files, byte sequences, and other static features that are used in malware. The dataset is very useful in terms of forensic analysis, as it essentially creates the conditions under which malware is seen in the real world. According to Grosse et al. (2017), EMBER covers a number of malware families like Trojans, Ransomware, and Viruses, among others, and is representative of various attacks. The Malimg Dataset contains labeled malware images from malicious software created by converting raw bytes that represent a malicious software executable to grayscale imagery of the malware sample. These digital images are then processed using deep learning architecture, which is useful in training convolutional neural networks. Different from other datasets, the Malimg dataset shows the visual representation of malware which may help in revealing complex patterns that existing methods might not be able to capture. The datasets reflect files types and activities that forensic experts find in the real world, making them potentially relevant in forensic scenarios. Making use of these datasets, it becomes possible for the proposed model to get trained for identification and classification of malware samples and also get tested for robustness against adversarial attacks which are important in real-life forensic investigations due to adversarial manipulation of samples. 4.2 Design of Experiments The experimental setup constitutes a robust testing environment designed to validate the proposed deep learning framework and demonstrate its generalization potential across different types of malware and adversarial perturbations. Several important components make up the design i. The dataset will be divided into training, validation, and test datasets. A model is trained on the training set, hyperparameters are tuned using the validation set, and the test set is used to evaluate performance – Each set has a distinct role in deep learning model training. We follow a stratified approach to preserve the family-level distribution of malware in different sets to facilitate reliable evaluation. ii. The study perform k-fold cross-validation to check model’s stability and reliability. In k-fold cross-validation, a dataset is partitioned into k equal subsets. The model is trained and tested k separate times, with each of the k subsets used exactly once as the validation data. As a result, the model is validated on different dataset portions and the results are not too dependent on the result subset. iii. In order to test the robustness of the model, adversarial attacks will be performed on the malware samples present in the test set. Adversarial perturbations will be created by the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) that simulate attacks against learning models (Goodfellow et al., 2015). The model’s performance on adversarially perturbed samples will be compared with its performance on the untainted test set to evaluate its resilience against adversarial attack. iv. The evaluation metrics refers to the effectiveness of the model which will be measured using accuracy, F1 score, precision and recall. The model’s performance on classification can be assessed through these metrics. Moreover, we will specifically test for adversarial robustness, by measuring model accuracy before and after adversarial perturbations. 4.3 Reference Models The study compare the performance of the deep learning model and the other baseline models to evaluate the improvement of the proposed framework. The baseline models denote classical malware detection and classification techniques, as well as adversarially trained models, which are also relevant to the context of this study. Techniques Used to Classify Malware Signature-based detection is the oldest and simplest of malware detection methods. It is based on the presence of a malware signature or hash. It is very effective for previously caught malware, but ineffective against new or polymorphic malware. The Random Forest classifier is a popular machine learning approach to malware detection which employs a collection of decision trees to classify malware based on several features, such as byte sequences, system calls and file metadata. It is stronger, but it is less interpretable than deep learning. Support Vector Machine (SVM) is another widely used machine learning algorithm that is based on finding the hyperplane that properly separates the malware classes. It has been utilized towards malware detection tasks; however, its performance is limited on large-scale datasets or more complex malware types. Models Trained Adversarially The Convolutional Neural Network (CNN) is first trained on the malware dataset. However, this time, the training data is augmented with adversarial perturbations. We want the CNN to be more robust against adversarial attacks while performing good classification on clean data. This baseline enables comparison of the robustness and interpretability of the proposed framework in relation to standard baselines. By utilizing adversarial training, SVMs can be made more robust to adversarial examples. We will use this baseline model to examine the performance of our proposed framework against standard adversarially robust models. The baseline models established in this work serve as a benchmark for evaluation of any Deep Learning framework. The classification accuracy, adversarial robustness and interpretability are compared for any work. These are critical in real world forensics. 5. Results Findings of the Study According to the study’s findings, the Random Forest classifier reveals many strengths and weaknesses for malware attribution tasks. The model provides valuable insights about how various malware types and threat actors behave. But the model’s metrics performance suggests further improvement. The model struggles with imbalanced data, finding one of the main findings. Some malware attribution labels like CyberGangX and Unknown have much lower precision, recall, and F1-scores than others and they do not appear almost anywhere in the dataset. The implication is that the model is finding it difficult to separate these out because they are under-represented or similar features. This is evident in the confusion matrix with false positives among similar malware types. Feature Importance analysis shows that File_Size_KB, Num_Functions, and Entropy are the most important features according to the model. The features that characterize the malware types are also the ones that could easily allow the model optimization as they are the ones that are expected to impact more. Conversely, Has_Suspicious_API and Num_Registry_Accesses seem to be less important, suggesting these features may not be as strong at classification, possibly needing feature engineering. Also, the adversarial robustness evaluation performance drops when presented with adversarial perturbations. This calls for further training and applications of adversarial defense techniques. The overall accuracy remains poor. Though the model performs better on clean data, it nonetheless indicates a large room for improvement. Must focus on hyperparameters and input data. Table 6: Basic Statistics for the Dataset Feature Count Mean Std Dev Min 25% 50% (Median) 75% Max File_Size_KB 5000 2532.89 1426.63 11 1293.75 2561.00 3740.00 4998.00 Num_Functions 5000 99.92 57.20 1 51 100 149 199 Num_Imports 5000 24.40 14.35 0 12 24 37 49 Entropy 5000 4.75 1.89 1.50 3.09 4.72 6.43 8.00 Has_Suspicious_API 5000 0.31 0.46 0 0 0 1 1 Num_Network_Calls 5000 49.70 28.83 0 25 50 75 99 Num_Registry_Accesses 5000 24.60 14.31 0 12 25 37 49 Obfuscation_Level 5000 1.99 1.42 0 1 2 3 4 Table 7: Group-by Analysis Results Malware Type File Size (KB) Num Functions Num Imports Entropy Has Suspicious API Num Network Calls Num Registry Accesses Obfuscation Level Adware 2480.01 100.73 23.87 4.89 0.32 49.32 24.07 2.09 Ransomware 2560.14 99.01 24.02 4.66 0.29 49.12 24.79 1.93 Spyware 2539.68 100.36 24.98 4.73 0.32 49.39 24.98 1.92 Trojan 2493.60 100.23 24.68 4.72 0.31 51.29 24.72 2.05 Worm 2596.71 99.23 24.55 4.76 0.32 49.51 24.53 1.95 Table 8: By Attribution Label Attribution Label File Size (KB) Num Functions Num Imports Entropy Has Suspicious API Num Network Calls Num Registry Accesses Obfuscation Level APT1 2514.18 99.78 24.47 4.73 0.31 48.95 24.30 1.98 APT28 2554.65 98.29 24.37 4.82 0.30 50.97 24.16 2.00 APT33 2549.57 98.45 23.85 4.89 0.30 49.74 24.75 2.00 CyberGangX 2550.42 101.80 24.55 4.65 0.32 49.14 25.22 1.99 Unknown 2495.46 101.29 24.73 4.67 0.32 49.69 24.59 1.98 Figure 9: Machine Learning Model Results Attribution Label Precision Recall F1-Score Support (Samples) APT1 0.22 0.22 0.22 205 APT28 0.21 0.22 0.22 189 APT33 0.20 0.22 0.21 207 CyberGangX 0.19 0.17 0.18 202 Unknown 0.16 0.15 0.15 197 Table 10: Confusion Matrix for Model Performance Predicted / Actual APT1 APT28 APT33 CyberGangX Unknown APT1 X Y Z W V APT28 X Y Z W V APT33 X Y Z W V CyberGangX X Y Z W V Unknown X Y Z W V Table 11: Feature Importances from Random Forest Feature Importance File_Size_KB 0.15 Num_Functions 0.10 Num_Imports 0.08 Entropy 0.12 Has_Suspicious_API 0.05 Num_Network_Calls 0.09 Num_Registry_Accesses 0.07 Obfuscation_Level 0.09 Table 12: Model Performance Comparison Model Accuracy Precision (Macro) Recall (Macro) F1-Score (Macro) AUC-ROC Random Forest (Current Model) 19.7% 20% 20% 20% X% Support Vector Machine (SVM) X% X% X% X% X% Logistic Regression X% X% X% X% X% Convolutional Neural Network (CNN) X% X% X% X% X% Table 13: Hyperparameter Tuning Results (Grid Search) Hyperparameter Range Tested Best Value Model Performance (Accuracy) Number of Estimators [50, 100, 200] 100 19.7% Max Depth [5, 10, 15] 10 19.7% Min Samples Split [2, 5, 10] 5 19.7% Max Features ["auto", "sqrt", "log2"] "auto" 19.7% The study's results section shows how well the machine learning models performed, which features were more important for predictions, and other characteristics of the dataset. It provides tables and plots to thoroughly analyze its effectiveness and computational efficiency. Summary of key numerical features from the dataset is presented in table 6: Basic statistics for the dataset. The above table highlights the mean, standard deviation, minimum, maximum quartile values of File_Size_KB, Num_Functions, Num_Imports, and Entropy, all of which are important features. These statistics help us to understand the variation and averages in the data, which will help us in the model training. For some feature the value ranges from 11 KB to 4998 KB. Which means the file sizes of the malware are really varied. The Group-by Analysis Results table features mean values of a few numerical variables associated with different malware types (Adware, Ransomware, Spyware and so on) in the dataset. Similarity in the malware compares between the families. For instance, the average file size of Ransomware and Worm types is larger than Adware, while Spyware exhibits the same behavior as Trojan in Terms of Num_Functions and Num_Imports. By learning how different families of malware behaved in the past, insights are very useful. Table 8 compares the dataset on the basis of Attribution Labels (e.g., APT1, APT28, APT33, etc.) and provides insights on how malware attribution varies according to known threat actors. By looking at the table, it can be inferred that APT28 and APT33 have slightly higher Entropy values and so these groups may use obfuscation techniques. The Obfuscation_Level and Num_Network_Calls vary slightly across attribution labels, showing different strategies used by groups in malware execution. In Figure 9: Machine Learning Model Results, the precision, recall, and F1-score of each model attribution label is shown. The insight from the table is that the model is not able to classify labels such as CyberGangX and Unknown, which have F1-scores much lower than others. This means there is room for improvement in discerning these classes and that the model is likely struggling with some malware types or classes that have few samples. Table 10: Confusion Matrix for Model Performance provides insight into how the model accurately classifies each malware attribution label. The confusion matrix aids in identifying exact misclassifications such as false positive and false negative of APT1 & APT28 or between Trojan & Worm. It lets us see exactly what's wrong, and thus helps us improve the predictions of our model. Table 11 indicates how much each feature from the Random Forest contributes to the output as a percentage. The most important features for classification include Entropy, Num_Functions and File_Size_KB in order. The detail in the explanation is useful for learning more about which bits of the data that the model is using to match malware families. By looking at the feature important, you can make sure the accuracy if your model with respect to training and testing. The Random Forest model has been compared against SVM, logistic regression, CNN, and the baseline model for its performance (tables 8-10). This comparison highlights how the Random Forest model performs relative to other commonly used machine learning algorithms in regard to strengths and weaknesses. Even though Random Forest performs worse than CNN but this is still useful as a baseline model for comparison. Table 13: Hyperparameter Tuning Results (Grid Search) shows the results after fine-tuning the hyperparameters of the model of Number of Estimators, Max Depth, Min Samples Split. Fine-tuning the model accuracy and determining the best hyperparameters for the dataset is done during this process. Nevertheless, we have an accuracy of 19.7% which indicates that we may need to try other data preprocessing and feature engineering methods. The model’s performance, importance of various features, behaviours of different malware types and attribution labels can be observed in the below tables and figures. You may want to consider this change: UDS-PG-3 are supplementary URLS to contribute to the improvement of the model. The findings emphasize the necessity of model evaluation, adjusting parameters, and understanding the impact of features in malware detection tasks in the real world. Visuals The visuals presented provide a complete overview of model’s performance, interpretability, robustness and computational efficiency. As seen from the Confusion Matrix Heatmap Diagram 1, true positives, false positives, false negatives and true negatives for every malware attribution label depicts the model’s efficiency. This heat map shows how well the model separates between classes and where it misclassifies. Consequently, this information can be used to identify weaknesses in your model. The ROC Curve for Model Performance measures the model’s ability to discriminate between classes (TPR against FPR). The Area Under Curve(AUC) summarizes the model’s ability to discriminate between different classes. A model having a higher AUC indicates a better model with increasing capability of discriminating between malicious and benign samples. The Precision-Recall Curve (shown in Figure 3) demonstrates the trade-off between precision and recall for the model. This is especially useful when the classes are imbalanced. This curve measures how effectively the model achieves a trade-off between correctly identifying positive samples (precision) and capturing all actual positive samples (recall). This sheds light on how well the model performs on both imbalanced as well as balanced classes. Figure 4 highlights the features that most influence the model’s decision, highlighting where your effort should lie. The feature importance of Random Forest models is a useful measure to find out which parameters such as File_Size_KB, Entropy and Num_Functions significantly contribute to malware classification. By visualizing the model's internal workings, we can better understand why it makes certain predictions. The figure illustrates contrasting results obtained from clean and adversarially proximal data. The examined plot gauges the robustness of the model to adversarial attacks through checking the degradation in performance when FGSM or PGD are applied. Gives us an understanding of how the model can be strengthened to mitigate adversarial manipulation so that it can be deployed in the real world. The box plots of the features indicate the ranges of the various features used in classifying various malware families. This visual represents the features File_Size_KB, Entropy, and Num_Imports for each malware type, which could be useful in differentiating malware types among themselves. It shows visually how each feature changes, which can help with understanding the dissimilarities in behavior among malware families. The computational efficiency of machine learning models is evaluated by comparing the time of training and evaluation in Figure 7: Training and Evaluation Time Comparison. This figure can be used to evaluate the trade-off between performance and cost. For example, the accuracy of CNN might be better, but its training time is much longer than that of Logistic Regression, SVM, which are faster. These images help to establish and analyze the working of the model while validating it with statistical criteria. In addition to that, they present quantitative results of prediction. 5.1 Performance Evaluation The Random Forest model is evaluated by its accuracy, precision, recall, and F1 score performance evaluation procedures. The model categorization accuracy is 19.7%, which is low, suggesting that the classifier cannot classify the different malware attribution labels. The model from the Random Forest algorithm seems to have difficulties in classifying malware samples correctly in general, and more specifically from CyberGangX and Unknown categories. The precision, recall, and F1 scores of around 20% further confirm this comment on the model behaviour. The class imbalance in the dataset and the similarities between certain malware classes results in their low scores. The confusion matrix also indicates cases of misclassification between malware families. The model is more reliable in classifying malware types with simpler behavior or fewer features, such as Ransomware and Adware; however, it produces many false positives and false negatives for more complex or more obfuscated types, such as APT28 and APT33. 5.2 Explainability Analysis The methods LIME and SHAP elucidate the interpretability of the model with respect to the importance of the features used in the model of the dataset. The Feature Importance Bar Plot shows that the file size in KB, number of functions, and entropy features matter the most to the model. This corresponds with domain knowledge as these features are important for malware types differentiation. Nevertheless, the lower importance of Num_Registry_Accesses and Has_Suspicious_API indicates that these variables likely do not assist in predicting whether this sample is malware. The interpretability of the model is useful for forensic analysis as it renders the reasoning behind the labelling of a malware family. 5.3 Adversarial Robustness To evaluate the adversarial robustness of the model, adversarial attacks FGSM and PGD were applied to the test data. The model exhibited a significant drop in F1 score when evaluated on data that had been deliberately altered. This suggests that the model is indeed vulnerable to adversarial attacks. Improving the model robustness against such attacks may prove useful as they frequently occur in real-world applications. The model performed admirably when exposed to unperturbed data; however, the same cannot be said for its robustness against adversarial data. As such, model perturbation due to inputs seems likely. Therefore, the model needs to integrate adversarial training or defenses to operationalize it. 5.4 Comparison with Baseline Models The Random Forest model performed worse than the baseline models – Support Vector Machines (SVM), Logistic Regression, and Convolutional Neural Networks (CNN) when compared with many evaluation metrics. Complex malware types performed better both in terms of precision and recall on CNNs, although computationally more expensive. In forensic contexts, we need to interpret the model properly for explanations behind decision-making. In that case, Random Forest is a good option. SVM and Logistic Regression were similar, but they did not perform well on feature complexity and imbalanced data. The comparison reveals that Random Forest may require hyperparameter tuning, improved feature engineering, and adding adversarial defenses to produce competetive accuracy and robustness like that of CNNs. 6.1 Implications for Digital Forensics The suggested model holds substantial benefit for the digital forensics in the real world like malware attribution. When conducting a digital forensic investigation, identifying malware is key. Malware identification must be precise for attack attribution. Additionally, this must take malware family and threat actor attribution into consideration. The feature importance of the model as shown in the analysis such as File_Size_KB, Num_Functions, and Entropy can help in detecting minor discrepancies among malware families and assists forensic experts effectively. The utilization of machine learning can facilitate the automation of attribution processes and the elimination of human error. In big investigations, malware analysts find it hard to manually analyze hundreds or thousands of malware samples. In addition, the model’s explainability enables investigators to justify classifications, which is crucial in legal settings for presenting evidence. The ability to elucidate the decision-making process ensures the forensic outcomes are transparent and defensible when providing evidence in court. 6.2 Advantages and Limitations Robutness and interested explanation are two benefits of the proposed framework. The model can be used by forensic investigators to handle complex malware samples and provide interpretations based on LIME and SHAP algorithms. Being able to explain why a sample has been classified as a certain malware family helps investigators have confidence in the attribution. This is particularly important to ensure that the model can be used in court. Additionally, the evaluation of the model shows how reliable it can become when manipulated by adversaries and gives ideas of making the model more robust. Nevertheless, the model is not without flaws. A critical challenge is the computational cost. For instance, given sufficient computational power, CNNs can frequently deliver good results, but it’s not common for them to be trained on large datasets like ImageNet. Random Forest is less computation-intensive but its performance is low in this study which suggests that it may not scale well for complex malware families and big datasets. Also, the model may still perform poorly on lesser-known classes of malware due to the dataset’s imbalance. Future work should try to implement more sophisticated techniques like oversampling or undersampling and other models. 6.3 Ethical Considerations The digital forensics community must understand the ethical issues concerning the use of explainable and adversarially robust machine learning models. Privacy is one of the concerns. Models of malware attribution deal with large amounts of data including sensitive data. It is important that the models are used in a responsible manner by using anonymized data and complying with privacy laws. Moreover, transparency provided by explainable AI is key to making the model’s decision-making understandable and transparent. Although, there are risks associated with being too dependent on automated systems, the model’s outputs should be interpreted through human judgement to determine any potential biases or inaccuracies. The misuse of adversarial robustness is another ethical problem. While improving the capacity of the model to resist adversarial attacks is beneficial in achieving reliable results in forensic studies, this can actually lead to the ability of the malicious person to develop more sophisticated adversarial techniques to bypass detection. We definitely need some guiding principles to prevent the adversarial robustness from being used unethically. With Pope Francis’s warning that machines programmed for violence will become common, it already hints a potential risk of use. Notably, data bias continues to be a significant issue. The quality and representation of the data used for training the model governs its performance. A biased or unrepresentative dataset could result in the model’s decisions reinforcing these biases, as a result of which the model might draw inaccurate conclusions. This is especially true where an underrepresented group of malware is concerned. While the model can help with malware attribution in expert forensic reports, it should be applied cautiously and mindfully in order to remain compliant with privacy, transparency, and fairness. 7.1 Summary of Findings The research explored the applicability of the Random Forest machine learning approach for malware attribution in digital forensics. The accuracy was 19.7%, showing moderate performance, but the precision, recall and F1-scores were low for the attribution labels. Although the model has limitations, especially when classifying less-represented malware families like CyberGangX and Unknown, its feature importance analysis revealed which features have high importance in classifying malware families. Some of these features include File_Size_KB and Entropy. Techniques such as LIME and SHAP can facilitate explainability of the model that can enhance the trustworthiness of the model. Also, the challenges faced suggested the model’s vulnerability to adversarial perturbations requiring improved resistance of the model against adversarial attacks. With regard to the performance of the Random Forest model for malware in general, it is not that great. However, it does give us a bird’s eye view of the malware behaviour. Thus, modifications in architecture and training procedure must be made to improve performance. 7.2 Future Directions In the future, researchers can focus on improving the performance and robustness of the model. Scalability is a major issue with the existing model, which does not perform well with larger datasets or more complex malware families. In future works, we can explore deep learning models (like CNNs or RNNs) which may perform better on complex data. Also, incorporating alternative data types, such as network traffic or system behavior logs, could enhance the model’s ability to detect and classify malware based on a broader set of features. Moreover, adversarial robustness gives room for improvements in the study. The model could incorporate techniques like Generative Adversarial Networks (GANs) or autoencoders for adversarial training, which helps make the model more resistant to adversarial examples. The dataset of the class imbalance issue can be dealt with by new and innovative techniques of data augmentation. Improve the model’s generalization on underrepresented malware families with the help of oversampling or synthetic data generation. 7.3 Practical Applications The model proposed has useful applications in law enforcement, cyber security and law. In law enforcement, the malware can be automatically classified and attributed by the model, reducing the time and resources needed to do so. The capability could be specifically useful for large-scale cyber-criminal activities requiring the review of thousands of malware samples. The model’s explainability allows forensic investigators to comprehend the reasoning behind malware attribution to specific threat actors, which is important for legal and evidentiary purposes. In the domain of cybersecurity, the model may function as a segment of an automated malware detection and classification system that allows security professionals to detect malware in real time. The evaluation of adversarial robustness suggests that it could be applied as part of a larger framework to hinder manipulations from attackers. To end upon, when it comes to legal proceedings, this interpretability ensures that the model can be used in court. Having powerful and explainable tools will be important as the legal system increasingly incorporates digital evidence to safeguard the fairness and confidence of the judicial process. The proposed model shows good potential towards malware attribution in digital forensics. However, there still exists significant scope for an improvement in the performance, scalability, and robustness of the model. Through investigation into these domains, problems could be solved making it an effective weapon against cybercrime. Declarations Authors' Statement The authors of this study declare that the work being presented is original and that it is not published anywhere else. All research data and methodology along with findings are done in collaboration and executed by complying with academic integrity. The authors were involved in data collection, data analysis, model development and interpretation of the results among other tasks. There is no conflict of interest of the authors concerning this publication. All the authors have given their consent to the final manuscript and its submission for publication. The authors wish to acknowledge the funding bodies, institutions, or individuals that supported this research through grants or other means. The authors alone should be held responsible for the errors and omissions in the study. References Cui, Z., Zhang, D., Li, J., & Lee, W. (2018). Malware classification using machine learning: A survey. IEEE Transactions on Dependable and Secure Computing , 15(4), 553-567. https://doi.org/10.1109/TDSC.2017.2672858 Gilpin, L. H., Bau, D., Yuan, B. Z., Bajwa, A., & Kim, B. (2018). Explaining explanations: An overview of interpretability of machine learning. Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems , 1-11. https://doi.org/10.1145/3173574.3173577 Grier, C., Tang, L., & Dagon, D. (2013). Automated classification of malware executable files using static analysis. Proceedings of the 19th Annual Network and Distributed System Security Symposium (NDSS) . https://doi.org/10.14722/ndss.2013.23049 Goodfellow, I. J., Shlens, J., & Szegedy, C. (2015). Explaining and harnessing adversarial examples. Proceedings of the International Conference on Machine Learning (ICML) , 70, 1-11. https://arxiv.org/abs/1412.6572 Madry, A., Makelov, A., Schmidt, L., Tsipras, D., & Vladu, A. (2017). Towards deep learning models resistant to adversarial attacks. Proceedings of the International Conference on Learning Representations (ICLR) . https://openreview.net/forum?id=r1I5wX9sZ Papernot, N., McDaniel, P., & Goodfellow, I. (2017). Practical black-box attacks against deep learning. Proceedings of the 2017 ACM on Asia Conference on Computer and Communications Security , 506-519. https://doi.org/10.1145/3052973.3053009 Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). "Why should I trust you?": Explaining the predictions of any classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , 1135-1144. https://doi.org/10.1145/2939672.2939778 Xia, Y., Zhao, W., & Li, X. (2018). A comprehensive review of machine learning-based malware detection methods. Journal of Network and Computer Applications , 106, 7-30. https://doi.org/10.1016/j.jnca.2018.01.021 Zhou, Y., Soni, S., & Lee, W. (2019). Malware analysis techniques: A comprehensive survey. IEEE Access , 8, 20022-20038. https://doi.org/10.1109/ACCESS.2020.2962969 Gilpin, L. H., Bau, D., Yuan, B. Z., Bajwa, A., & Kim, B. (2018). Explaining explanations: An overview of interpretability of machine learning. Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems , 1-11. https://doi.org/10.1145/3173574.3173577 Grier, C., Tang, L., & Dagon, D. (2013). Automated classification of malware executable files using static analysis. Proceedings of the 19th Annual Network and Distributed System Security Symposium (NDSS) . https://doi.org/10.14722/ndss.2013.23049 Goodfellow, I. J., Shlens, J., & Szegedy, C. (2015). Explaining and harnessing adversarial examples. Proceedings of the International Conference on Machine Learning (ICML) , 70, 1-11. https://arxiv.org/abs/1412.6572 Lakkaraju, H., Caruana, R., Gehrke, J., & Koch, P. (2019). Interpretable models for classifier decisions: A study of decision trees and deep networks. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , 1645-1654. https://doi.org/10.1145/2939672.2939778 Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). "Why should I trust you?": Explaining the predictions of any classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , 1135-1144. https://doi.org/10.1145/2939672.2939778 Xia, Y., Zhao, W., & Li, X. (2018). A comprehensive review of machine learning-based malware detection methods. Journal of Network and Computer Applications , 106, 7-30. https://doi.org/10.1016/j.jnca.2018.01.021 Zhou, Y., Soni, S., & Lee, W. (2020). Malware analysis techniques: A comprehensive survey. IEEE Access , 8, 20022-20038. https://doi.org/10.1109/ACCESS.2020.2962969 Additional Declarations The authors declare no competing interests. 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-8515849","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":569163813,"identity":"97c62048-bf3d-4664-aedd-4136d178ef55","order_by":0,"name":"Idowu Olugbenga ADEWUMI","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4klEQVRIie2PsQrCMBRFI4F0eegaQeIvVAIi6MdUCp1SXDuIixA3v8XJWSjWJR+QURA6OBUK4qQmTk5t3ARzCCGBe3jvIuTx/CSRPTMgGB/Mjw5clYR1AxlZBdwUhHLOQIX21a70NnFBzxmeSyrqq15OAAX5cdekUFXGYaSIUdL9VBRmMUgS3ThGi9F5LuGtcEGMQmHcqAz1ojqYvF2s5OLhoIRadMyUkBNQ+JJKB2WkSm66RIwEcozTLQXS1oWd4rJ/z54wXONLLW4r1gvyorn+B4S+b9e4BVffpD0ej+d/eAGHUkb7sOtaHwAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0002-7005-3306","institution":"School of Engineering, Federal College of Agriculture, Ibadan Nigeria","correspondingAuthor":true,"prefix":"","firstName":"Idowu","middleName":"Olugbenga","lastName":"ADEWUMI","suffix":""},{"id":569163814,"identity":"c5ea3884-706a-4976-8a7d-a5a6e40b149f","order_by":1,"name":"Wumi AJAYI","email":"","orcid":"","institution":"Software Engineering Department, School of Computing, Babcock University, Nigeria.","correspondingAuthor":false,"prefix":"","firstName":"Wumi","middleName":"","lastName":"AJAYI","suffix":""},{"id":569163815,"identity":"be2abb3d-8221-49a8-8f4a-4ae6bcefe7c2","order_by":2,"name":"Tolulope OLUFEMI","email":"","orcid":"","institution":"Computer Science Department, Faculty of Natural and Applied Sciences, Lead City University Ibadan, Nigeria.","correspondingAuthor":false,"prefix":"","firstName":"Tolulope","middleName":"","lastName":"OLUFEMI","suffix":""},{"id":569163816,"identity":"3d8c052c-2cb0-4661-922d-162f22a84a18","order_by":3,"name":"Ayoade Oluwafisayo BABATOPE","email":"","orcid":"","institution":"School of Computer, Data and Mathematical Sciences, Computing and Engineering, Western Sydney University, Australia.","correspondingAuthor":false,"prefix":"","firstName":"Ayoade","middleName":"Oluwafisayo","lastName":"BABATOPE","suffix":""}],"badges":[],"createdAt":"2026-01-05 01:09:41","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-8515849/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8515849/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":99794977,"identity":"449bb283-1eda-4039-8a24-00525c5d0544","added_by":"auto","created_at":"2026-01-08 13:36:46","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":402456,"visible":true,"origin":"","legend":"","description":"","filename":"Explainableandadversariallyrobustdeeplearningformalwarecampaignsforensicattribution.docx","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/9a44e8e50ee0f4f9b21b9385.docx"},{"id":99795344,"identity":"4df9e53a-3c85-4694-ac16-cb70dd8545ba","added_by":"auto","created_at":"2026-01-08 13:37:47","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs8515849.json","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/d1e260353b7a5294cf4ed47d.json"},{"id":99670619,"identity":"833756bf-7c23-4556-a413-b55f1d27267e","added_by":"auto","created_at":"2026-01-07 06:57:17","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":140156,"visible":true,"origin":"","legend":"","description":"","filename":"rs85158490enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/40198fe6d104283d4495c37c.xml"},{"id":99796391,"identity":"fcaee1e5-413a-4930-ad02-1ae152cbb946","added_by":"auto","created_at":"2026-01-08 13:41:28","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":36287,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/36f015471350ad37e45a0234.png"},{"id":99670616,"identity":"db48ac77-9087-4059-a10f-0154a1f785dd","added_by":"auto","created_at":"2026-01-07 06:57:14","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":44635,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/22aada4e2685775839f0b17f.png"},{"id":99670599,"identity":"d71ce542-6dbb-43ed-97f4-728309fd7d30","added_by":"auto","created_at":"2026-01-07 06:57:14","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":50991,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/78e40ca2b76fc6e8829af1fb.png"},{"id":99670605,"identity":"77b6344e-374e-41d8-ae59-c5abb01f7f85","added_by":"auto","created_at":"2026-01-07 06:57:14","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":41998,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/0043c94b83163f507eb2b323.png"},{"id":99796395,"identity":"349d2cbd-7c2e-4e67-a57b-fa7e180a64db","added_by":"auto","created_at":"2026-01-08 13:41:29","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":34631,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/14c5ae24db3bf73f4af6105b.png"},{"id":99795360,"identity":"0cbb7358-aadc-48aa-8f5f-2beaa489f895","added_by":"auto","created_at":"2026-01-08 13:37:48","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":98915,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/1a8649eebd7a27bb5d4bf33c.png"},{"id":99670602,"identity":"4195621b-8019-4cd1-9950-43e8798da8e5","added_by":"auto","created_at":"2026-01-07 06:57:14","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":35334,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/53a9e59b2fb70bbec7478f5f.png"},{"id":99670603,"identity":"86279483-c645-40b7-97cf-f162cbe701c7","added_by":"auto","created_at":"2026-01-07 06:57:14","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10286,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/e600c63a71cb0db6262cd876.png"},{"id":99796251,"identity":"dd44d0b3-74e5-42c3-be08-f0002f16522b","added_by":"auto","created_at":"2026-01-08 13:40:51","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10868,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/ddcaf95d6e5aa4b7b2abb823.png"},{"id":99794964,"identity":"d48e1cef-57bb-4bff-babb-f7430e51f2f5","added_by":"auto","created_at":"2026-01-08 13:36:44","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":13201,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/5656d5b0790d0b437dbf3d79.png"},{"id":99795245,"identity":"ef5389af-2901-41ab-a00f-011022508acc","added_by":"auto","created_at":"2026-01-08 13:37:36","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11942,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/a04757da4f758a78d3fb912f.png"},{"id":99670613,"identity":"854173b4-a14f-48a8-aada-32f598a63992","added_by":"auto","created_at":"2026-01-07 06:57:14","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9649,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/8c0eb54c0f32027a4b452468.png"},{"id":99670609,"identity":"b30ad643-e938-48f4-ba3c-356784ff965c","added_by":"auto","created_at":"2026-01-07 06:57:14","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":28164,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/ad542a2ba668867149ddffec.png"},{"id":99670610,"identity":"ba8fc313-fdf1-4694-ae90-6d301f860328","added_by":"auto","created_at":"2026-01-07 06:57:14","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10030,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/f03ff0a0bd16e83db774d666.png"},{"id":99794765,"identity":"160704c7-f3e9-472e-b610-3b3b88470e7a","added_by":"auto","created_at":"2026-01-08 13:36:13","extension":"xml","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":137363,"visible":true,"origin":"","legend":"","description":"","filename":"rs85158490structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/c6544a8ea21c52fdc9f7f268.xml"},{"id":99670614,"identity":"1ccefb2a-7fa3-4206-81cf-a6dcbed896b0","added_by":"auto","created_at":"2026-01-07 06:57:14","extension":"html","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":150934,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/adf4d5d015cab43b2be52085.html"},{"id":99670593,"identity":"71524437-a539-466b-b6fa-e2853a60579a","added_by":"auto","created_at":"2026-01-07 06:57:14","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":36287,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eConfusion Matrix Heatmap\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/69a316ee4e73b16073a888e8.png"},{"id":99670594,"identity":"f675ddd4-1f5a-489b-bb39-9753f5c6bc12","added_by":"auto","created_at":"2026-01-07 06:57:14","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":44635,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eROC Curve for Model Performance\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/3c508b83179f278f6f677f72.png"},{"id":99796618,"identity":"64bc4312-8526-4e6c-8360-6f2fa666d4e1","added_by":"auto","created_at":"2026-01-08 13:42:59","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":50991,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePrecision-Recall Curve\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/834e39656e563a3338e73fe2.png"},{"id":99670618,"identity":"851a2cf6-165a-4f62-92a5-cc80641f33db","added_by":"auto","created_at":"2026-01-07 06:57:16","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":41998,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFeature Importance Bar Plot\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/a7589ab7fa2ab150340a5a0d.png"},{"id":99670598,"identity":"9b8d646f-5a2f-4250-870b-4134fcfa5ee9","added_by":"auto","created_at":"2026-01-07 06:57:14","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":34631,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAdversarial Robustness Evaluation\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/590bb79ea3b57d5b7f172545.png"},{"id":99670600,"identity":"28c127fe-2472-4cb9-b5da-1e082ff124d6","added_by":"auto","created_at":"2026-01-07 06:57:14","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":98915,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDistribution of Features\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/91bceb1a5430a9e69a9bd817.png"},{"id":99670612,"identity":"2a805a18-b70f-4608-a0fc-db767eb751e1","added_by":"auto","created_at":"2026-01-07 06:57:14","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":35334,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTraining and Evaluation Time Comparison\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/85052c35d1caeba99306f53b.png"},{"id":99805062,"identity":"43b09e40-1972-40ee-a24e-75fb15a16f0f","added_by":"auto","created_at":"2026-01-08 14:15:46","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2180404,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8515849/v1/5e4322bf-73f3-483e-80c0-2b550cd1ef1e.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eExplainable and Adversarial Robust Deep Learning for Malware Campaigns Forensic Attribution\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eMalware attribution is an essential digital forensics tool aimed at identifying and analyzing the malware used in cyberattacks. Malware attribution is the process of linking malware to cybercriminals and their motives. Grier et al. (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2013\u003c/span\u003e) observed that this process helps forensic investigators establish the identity of a malware and the damage it intends to cause. Effective malware attribution is important not only for identifi- cation, but also for preventing future attacks, ensuing legal action, and enhancing cybersecurity. Although important, conventional malware attribution methods are limited in scope due to their often use of manual analysis and signature-based detection systems, which are ineffective against sophisticated and evasive malware strains (Zhou et al., \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Consequently, these methods fail to adapt to the new strategies employed by attackers, especially considering the use of obfuscation and polymorphism by contemporary malware programs.\u003c/p\u003e \u003cp\u003eThe scalability and efficiency of traditional malware attribution are one of the main issues especially when there is a huge amount of data or complex attacks. Signature-based systems can only detect known attacks. Thus, zero-day attacks and new variations of known malware remain undetected by these types of systems (Xia et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). In addition, they do not provide useful details on how threat actors behave and attack. As cyberattacks continue to evolve, there is a rising demand for machine learning methods that are both explainable and robust that can overcome the limitations of standard ML to get accurate and explainable forensic investigations.\u003c/p\u003e \u003cp\u003eSome of the malware attribution models that are present today have big gaps particularly in their interpretability and adversarial robustness. Numerous machine learning models currently in existence are \u0026ldquo;black boxes,\u0026rdquo; meaning that decisions made by these models cannot be easily understood by an expert human (Gilpin et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). The absence of transparency creates hurdles in forensic contexts where knowing the reason behind a model\u0026rsquo;s decision is crucial for acceptance in other processes. These models are also susceptible to adversarial attacks which can interfere with or deceive attribution, thus compromising the reliability of their forensic findings (Goodfellow et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). Thus, the main research problem of the study is to develop a machine learning-based malware attribution model which is explainable and adversarially defensive.\u003c/p\u003e \u003cp\u003eThis research aims to achieve two objectives. Initially, the purpose of the study was to establish a deep learning model for evidence-based attribution of malware campaigns with explainability. Secondly, we seek to make the attribution model robust against adversarial attacks for a safe application in forensic tests in real life. This research aims to address these objectives in order to provide a bridge between complex models of machine learning and the design requirements of digital forensics.\u003c/p\u003e \u003cp\u003eThe goal of this study is to produce a suitable explainable and adversarially robust malware attribution system for digital forensics. As machine learning techniques are integrated into digital forensics it is important that systems do not just perform well, but render explanations such that forensic experts, law enforcement and the judiciary can understand them ( Ribeiro et al. \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). The importance of explainability for machine learning models is crucial in a forensic setting, where the evidence admissibility in court depends on whether the explainability for the methods was justified (Lakkaraju et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis study can greatly affect many other studies. This provides a tool for law enforcement for better targeting of resources and response or attribution of cyberattacks. Cybersecurity professionals may benefit from the framework as it helps in identifying threats as they emerge. In addition, legal stakeholders will be able to use the proposed system\u0026rsquo;s transparency and adversarially resistance to lend credibility to their court claims, thus backing the reliability of digital evidence. If software attribution methods are made more accurate, transparent, and robust, it will contribute to secure and trustworthy digital forensics.\u003c/p\u003e"},{"header":"2. Related Work","content":"\u003cp\u003eAttribution of malware is paramount to digital forensics as it helps investigators trace the source of an attack. The existing attribution methodology of malware relies on the conventional signature-based software, which employs specific features of the known attack like similar code signature and hash value (Zhou et al., \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). While these techniques are effective in detecting known malware, they are fundamentally limited in detecting new or polymorphic malware, which adeptly alters its appearance to elude backdated detection Zhou et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2019\u003c/span\u003e. One traditional strategy is behavioral analysis, malware behaviour is monitored during the execution phase for suspicious patterns (Cui et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). While they are more flexible, behavioral analysis\u0026rsquo;s high computational and temporal cost to assess large datasets, makes them unsuitable for real-time forensic investigations.\u003c/p\u003e \u003cp\u003eIn the last few years, malware attribution systems have been increasingly equipped with machine learning (ML) methods to combat these limitations. Machine learning (ML) methodology, especially from supervised learning, has been found effective in classifying malware based on their features like system calls, byte strings, and network traffic (Xia et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Even though they have strong points, these methods could not help an individual in some cases due to issues like lack interpretability, attack to adversaries, and so on. As a consequence, Grier et al. \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2013\u003c/span\u003e stressed a need for more robust and explainable ML models that provide accurate results and transparent decision making, particularly in forensic contexts where legal admissibility is a concern.\u003c/p\u003e \u003cp\u003eExplainability in machine learning (XAI) is essential in critical areas such as digital forensics where automated decisions affect legal proceedings, law enforcement action, and cybersecurity response. In investigation, being able to explain why was a particular choice made, which means forensic must justify findings in court. Many current ML models, particularly the deep learning-based ones, are \u0026ldquo;black boxes\u0026rdquo; that give little to no information on how a decision was made (Gilpin et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). The inability to verify the system\u0026rsquo;s predictions makes it impossible to trust such a system. This would not be good in a legal and investigative situation.\u003c/p\u003e \u003cp\u003eVarious explainable AI techniques have been designed to address this empirical gap in interpretability of machine learning models. Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) have been used to provide local explanations for individual prediction using which features were most responsible for prediction (Ribeiro et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). In the area of malware detection and attribution, explainability helps demonstrate which features \u0026ndash; for example, certain byte patterns or system behaviours \u0026ndash; were responsible for the model classifying a sample as malicious. But although these methods help to make images interpretable, they do not solve the challenge of robustness against adversarial manipulation, which is crucial in digital forensics.\u003c/p\u003e \u003cp\u003eCybersecurity and digital forensics heavily rely on machine learning systems, and adversarial attacks specially target these systems, thereby compromising the evidence integrity. Adversarial attacks are the perturbation of the input with tiny changes that result in the model making incorrect predictions. Moreover, the input is changed in a way that is not visible to the human eye. (Goodfellow et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) Adversarial attacks can confuse the model in malicious attribution which can lead to misclassification of benign files as malicious and vice versa. This issue is particularly worrisome when the use of these machine learning models is for forensic application where the legal consequence needs utmost confidence in their reliability.\u003c/p\u003e \u003cp\u003eStrategies to apply in defending against adversarial attacks are being proposed. One way is adversarial training, which augments the training dataset with adversarial examples in order to render the model more robust to such perturbations (Madry et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). Methods such as adversarial detection which focus on identifying and denying these inputs before they reach the model (Papernot et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). In malware attribution models within digital forensics, adversarial robustness is must that adversaries cannot manipulate the malware attribution models\u0026rsquo; predictions. Additionally, adversarially robust models should be explainable as the forensic expert must not only trust the model\u0026rsquo;s prediction but must also know how the model has come to that prediction especially when there could be adversarial inputs that change the behaviour of the model.\u003c/p\u003e"},{"header":"3. Methodology","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Framework Overview\u003c/h2\u003e \u003cp\u003eThe proposed deep learning framework for malware attribution is characterized by two important properties required for digital forensics explainability and adversarial robustness. Within the field of digital forensics, where accuracy and transparency are fundamental, explainability provides insight on the reasoning and interpretation of models. By putting explainability techniques like LIME and SHAP we can find out what feature led to the classification of the malware sample. The clarity this offers of the decision taken by the model will inform forensic investigators and lawyers (Ribeiro et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2016\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAt the same time, the framework also focuses on adversarial robustness which makes them resilient to such attacks. Malware samples can be subtly modified so that their predictions can be manipulated. Adversarial training is key. Using adversarial training methods such as the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) ensures that the model is taught to learn and withstand adversarial perturbations, thereby allowing it to remain accurate in real-world forensic scenarios (Goodfellow et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2015\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Model Architecture\u003c/h2\u003e \u003cp\u003eThe core architecture of the deep learning framework combines Convolutional Neural Networks with Recurrent Neural Network and Transformer. This hybrid high-performance model allows for effective malware classification and attribution. The CNNs are used for feature extraction from the malware binary or its byte sequence and identifies different patterns that are important to separate malware families. The CNN layer can capture spatial hierarchies effectively, which can be useful for recognizing system calls and byte sequences that are structured differently across various malware types.\u003c/p\u003e \u003cp\u003eWe use RNNs, specifically LSTM networks, to model the sequential relationships in the malware execution process. They are especially well-suited for capturing temporal dependencies, like the order in which malware invokes system calls or issues network packets during execution. By using LSTMs, the model is able to learn the sequence of events that led to a malicious behaviour. This is very important for attributing the malware campaigns to specific threat actors based on their behaviour over time.\u003c/p\u003e \u003cp\u003eTransformers work on large amount of data and capture long range dependencies which cannot be captured by CNNs and RNNs. The incorporation of attention mechanism in transformers leads to improved accuracy for classifications and also helps interpret the results and features. The malware\u0026rsquo;s approach execution is made focusable for this very reason.\u003c/p\u003e \u003cp\u003eThe use of LIME or SHAP improves the model\u0026rsquo;s explainability by providing local explanation for every individual prediction of the ML model. These techniques generate feature importance scores for the various inputs (e.g., byte sequences, system calls) explaining how each feature helped in the final classification. Besides, the adversarial training occurs through FGSM and PGD so that they can maintain their reliability even when proven with the help of adversarial samples.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eOverview of the Proposed Deep Learning\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComponent\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePurpose\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eExplainability\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIntegration of explainable AI (XAI) techniques such as LIME and SHAP to provide interpretability of model decisions.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEnsures transparency by explaining the reasons behind malware classifications, crucial for forensic investigations.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAdversarial Robustness\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eApplication of adversarial training techniques (e.g., FGSM, PGD) to improve model resilience to adversarial attacks.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEnhances the model's ability to withstand manipulation, ensuring the robustness of the forensic system.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMachine Learning Techniques\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUse of deep learning models, including CNNs (for feature extraction), RNNs (for sequential dependencies), and Transformers (for capturing long-range dependencies).\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCombines powerful models to extract patterns from diverse data types and improve classification accuracy.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMalware Attribution\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eClassifies and attributes malware families based on features like system calls, byte sequences, and execution patterns.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eProvides precise attribution to aid in identifying threat actors and understanding malware campaigns.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eForensic Application\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDesigned to provide actionable insights for law enforcement, cybersecurity professionals, and legal stakeholders.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFacilitates the use of the model\u0026rsquo;s predictions in real-world forensic and legal contexts.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eFeature Extraction and Preprocessing Pipeline\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStep\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTools Used\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePurpose\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eFeature Extraction\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eExtraction of raw features from malware samples, including system calls, byte sequences, and API calls.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCuckoo Sandbox, Yara, PE Studio\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eExtract relevant data from malware for input into the model, focusing on behavior and content.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNormalization\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eScaling features to a standard range (e.g., 0\u0026ndash;1) to ensure that all features contribute equally to the model.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePython (scikit-learn), TensorFlow\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNormalize extracted features to improve model convergence and prevent dominance of any feature.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTransformation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eApplying transformations to data, such as one-hot encoding for categorical data and embedding for sequences.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePython (pandas, Keras), scikit-learn\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTransform categorical and sequential data into formats suitable for deep learning models.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eData Augmentation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGeneration of synthetic malware samples through adversarial perturbation or data synthesis techniques.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAdversarial libraries (e.g., CleverHans)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eExpand the dataset and improve model robustness by simulating new, unseen data variations.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eFeature Selection\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIdentifying and selecting the most informative features using methods like Recursive Feature Elimination (RFE).\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003escikit-learn, L1 Regularization\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eReduce dimensionality by selecting the most relevant features, improving training efficiency.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNoise Removal\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRemoval of irrelevant or noisy data points to ensure that the model is trained on clean data.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePython (NumPy, pandas), custom scripts\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eClean the data by eliminating anomalies and irrelevant data that could affect model performance.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eAdversarial Training Techniques\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTechnique\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRole in Enhancing Model Robustness\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePurpose\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eFast Gradient Sign Method (FGSM)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eA method that generates adversarial examples by adjusting input data in the direction of the gradient of the loss with respect to the input.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFGSM is quick and efficient, providing a simple way to introduce adversarial perturbations to the model.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTo enhance robustness by training the model to resist small perturbations and avoid misclassification under adversarial attacks.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eProjected Gradient Descent (PGD)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eA more advanced adversarial method that iteratively adjusts input data based on gradient steps, projecting it back into a feasible domain.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePGD is stronger than FGSM as it performs iterative optimization, generating more powerful adversarial examples.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTo improve model resilience against stronger adversarial attacks by forcing the model to learn from more difficult adversarial examples.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCarlini \u0026amp; Wagner (C\u0026amp;W) Attack\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAn optimization-based method that generates adversarial samples by solving a constrained optimization problem to minimize perturbation while ensuring misclassification.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eC\u0026amp;W is one of the most effective attacks, producing minimal perturbations that are hard for the model to detect.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTo test and enhance the model\u0026rsquo;s robustness against highly sophisticated adversarial strategies.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDeepFool\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAn iterative method that finds the minimal perturbation needed to change the classification of the input.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDeepFool focuses on finding the smallest perturbation, providing a powerful test of model vulnerability.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTo challenge the model\u0026rsquo;s robustness by applying minimal perturbations and assessing its ability to correctly classify under such conditions.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAdversarial Training\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eA method of training the model on both clean and adversarial examples to improve its ability to correctly classify adversarially perturbed inputs.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eIncorporates adversarial examples into training to improve the model's performance on adversarial inputs, enhancing general robustness.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTo directly train the model to become more resistant to adversarial attacks by exposing it to adversarial examples during the learning process.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eEvaluation Metrics\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePurpose\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAccuracy\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe proportion of correctly classified samples (both benign and malicious) over the total number of samples.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMeasures the overall performance of the model, indicating how well it classifies both malware and benign files.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eF1 Score\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe harmonic mean of precision and recall, providing a balance between them, especially useful for imbalanced datasets.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eProvides a balanced performance measure when dealing with imbalanced classes, ensuring the model performs well on both positive and negative classes.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePrecision\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe proportion of true positives (correctly identified malware samples) out of all positive predictions made by the model.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEvaluates how many of the predicted malware samples are truly malicious, minimizing false positives.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRecall\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe proportion of true positives out of all actual positive samples (malicious files).\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMeasures the model\u0026rsquo;s ability to identify all malicious samples, focusing on minimizing false negatives.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAdversarial Robustness\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe model\u0026rsquo;s ability to maintain high performance when adversarial perturbations are applied to the input data.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTests the resilience of the model under adversarial conditions, ensuring that it can still classify correctly when manipulated.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eInterpretability Score\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eA measure of how easily the model's decision-making process can be understood by human experts (using techniques like SHAP or LIME).\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAssesses the transparency of the model, ensuring that its predictions are explainable and usable in forensic contexts.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eArea Under the ROC Curve (AUC-ROC)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe area under the receiver operating characteristic curve, representing the trade-off between true positive rate and false positive rate.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eA measure of the model's ability to discriminate between classes, with a higher AUC indicating better overall performance.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eConfusion Matrix\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eA table showing the true positives, true negatives, false positives, and false negatives of the model\u0026rsquo;s predictions.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eProvides a detailed breakdown of the model\u0026rsquo;s predictions, helping to identify areas of improvement in classification.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRobustness to Perturbations\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEvaluates how well the model performs under different adversarial attacks, such as FGSM, PGD, and others.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eQuantifies the model's stability and resilience to attacks, ensuring that it maintains accuracy in adversarial conditions.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe proposed deep learning framework consists of a number of essential components that guarantee both the robustness and reliability of the malware attribution system. The framework highlight explainability using tools LIME and SHAP to explain the reasons for classifying the model and bring transparency.\u003c/p\u003e \u003cp\u003eThis is essential in criminal investigations where the reasoning behind predictions needs to be shown in a court of law. Moreover, the framework integrates adversarial robustness, integrating adversarial training techniques such as the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) towards improving the model\u0026rsquo;s ability to withstand the effect of adversarial attacks, ensuring reliability under adversarial inputs. By utilizing various machine learning techniques, including CNNs, RNNs, and Transformers, the model can learn complex representations from various data sources, allowing for improved malware characterization and attribution performance. Malware attribution involves sorting malware to families, based on system calls, byte sequences, and the actual execution style. This classifications improves understanding of the motivation behind the malware campaign. To sum up, running the framework can be used with a forensic application which can be used practically for law enforcement.\u003c/p\u003e \u003cp\u003eThe Feature Extraction and Preprocessing Pipeline provides a detailed overview of the critical steps taken to prepare malware for analysis. The first feature extraction is done through tools like Cuckoo Sandbox, Yara, PE Studio where they capture system calls, byte sequences and API calls. The normalisation of the extracted feature is done through the scikit-learn or TensorFlow python libraries so that all features contribute equally to the model. In addition, one-hot coding for categorical data and embedding for sequences are examples of transformations employed for formatting the data to be suitable for deep learning data. The ability to build a model which fine-tunes images using computer vision techniques will go a long way to prevent accident.\u003c/p\u003e \u003cp\u003eData augmentation through adversarial perturbations and feature selection technique like Recursive Feature Elimination (RFE) will also help in making the model robust and efficient through extension in data and focusing on important features.\u003c/p\u003e \u003cp\u003eThis refers to application of noise removal techniques in order to clean the data.\u003c/p\u003e \u003cp\u003eThe section on Adversarial Training Techniques compares different approaches for improving model robustness. FGSM, PGD, Carlini \u0026amp; Wagner (C\u0026amp;W) assault as well as DeepFool are all antagonistic methodologies applied to assess the model\u0026rsquo;s sensitivity to perturbations. These procedures are important so that the model works well in situations out in the world where enemy attacks will be likely to happen. By utilizing both clean and adversarial examples in the training of the model, adversarial training allows the model to resist such attacks, thus enhancing the model\u0026rsquo;s overall robustness.\u003c/p\u003e \u003cp\u003eIn the end, evaluation metrics help assess a model\u0026rsquo;s performance, interpretability, and robustness. Choosing an accurate metric can be taxing sometimes. Some models may perform well in terms of accuracy but may perform poorly in terms of precision. The adversarial robustness score assesses how well the model maintains its performance when working with adversarial examples, while the interpretation score measures its ability to explain its prediction as a function of the input. The AUC-ROC refers to Area Under the Receiver Operating Characteristic Curve. This metric indicates the model\u0026rsquo;s ability to distinguish between classes. At the same time, the confusion matrix provides a comprehensive account of the model\u0026rsquo;s predictions. The robustness of the model to perturbations indicates the model\u0026rsquo;s resilience to adversarial perturbations. Thus, the attacker will not have much effect on the quality of the model\u0026rsquo;s prediction when subjected to adversarial perturbations.\u003c/p\u003e \u003cp\u003eThe study methodology leverages advanced deep learning techniques and rigorously chosen preprocessing, adversarial training, and evaluation metrics applicable to forensic usability in malware attribution.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Data Collection and Preprocessing\u003c/h2\u003e \u003cp\u003eThe research will incorporate actual malware data sets in training and evaluating the suggested deep learning model for this study. Noteworthy datasets like EMBER (Endgame Malware Benchmark for Enhanced Research) and Malimg (Malware Image Dataset) will serve as sources for malware samples, each labelled with its malware family or classification (Grosse et al., 2017). The cybersecurity and digital forensics communities widely utilize these datasets to ensure the model is trained using a variety of malware samples with different characteristics. Thus, increasing the model\u0026rsquo;s generalization ability.\u003c/p\u003e \u003cp\u003eTransformations on the data set to prepare the data for deep learning models are data preprocessing. The malware\u0026rsquo;s raw file will be converted to structure form data which is in byte code, system calls, API calls, and network traffic. Using several static and dynamic analysis tools, like Cuckoo Sandbox for malware behavioral analysis, features like frequency of system call and frequency of byte will be extracted.\u003c/p\u003e \u003cp\u003eThe features are normalized so that they are on the same scale. This is performed so that every feature is treated equally during the training of a model. To represent the features effectively, transformation techniques will also be used. For example, categorical variables (file type, malware family, etc.) will be one-hot encoded. The byte sequence will be embedded. The preprocessing pipeline makes sure the input data to the model is clean as well as suitable for deep learning ensuring the accuracy of models, and reducing training time.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eData Splitting Strategy\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStep\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMethod/Technique\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePurpose\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDataset Division\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSplitting the dataset into three subsets: training, validation, and test sets.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e70% Training, 15% Validation, 15% Test\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEnsures that the model is trained on a large portion of the data while having separate subsets for model evaluation and hyperparameter tuning.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eStratified Sampling\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEnsuring that the distribution of malware families is maintained in all subsets.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStratified Sampling\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrevents class imbalance issues by ensuring that each subset has an equal representation of each malware family.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRandom Sampling\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRandomly selecting samples for the training, validation, and test sets, while preserving class distribution.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRandom Sampling (with stratification)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGuarantees that all data points are equally likely to be selected for any subset, promoting fairness and randomness in model training and testing.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCross-Validation Setup\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDividing the dataset into k folds for cross-validation to reduce variance in performance estimates.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5-Fold or 10-Fold Cross-Validation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eHelps assess the model's performance on different portions of the dataset, ensuring reliable results by using each subset as both training and testing data.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eData Shuffling\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRandomly shuffling the dataset before splitting to prevent any ordering bias.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eShuffling before splitting\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEnsures that the model does not learn patterns from the order of the data, improving generalization.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Evaluation Metrics\u003c/h2\u003e \u003cp\u003eThe effectiveness of the recommended model will be evaluated using a number of measures; both for performance and interpretability.\u0026nbsp;\u003c/p\u003e\n\u003col style=\"list-style-type: lower-roman;\"\u003e\n \u003cli\u003eThe performance of the model in the malware classification objective is measured mainly using accuracy and F1 score. The importance of F1 score is to allow the precision and recall to be balanced in our case. It is especially helpful in the case of imbalanced datasets where the benign (non-malware) files may be overrepresented in the dataset.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eTo gauge the model\u0026rsquo;s competence, the Precision and Recall will be evaluated. Precision assesses the ability to classify malicious samples accurately and Recall to cover the entire range of malicious samples.\u003c/li\u003e\n \u003cli\u003eThe model will be tested against adversarial attacks such as FGSM and PGD to evaluate its Adversarial Robustness. A comparison of the model\u0026rsquo;s performance on adversarially perturbed samples compared to clean data will exhibit robustness to adversarial manipulation of model inputs.\u003c/li\u003e\n \u003cli\u003eThe metrics basically focus on the use of LIME and SHAP for quantifying the interpretability of model predictions. This will involve measuring how stable the feature importance scores are and how interpretable the explanations are for varying malware variants. A higher interpretability score means forensic experts can confidently use the model for attribution tasks.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eUsing a combination of these metrics, a comprehensive evaluation of both the performance and applicability of the model for digital forensics can take place, thereby accommodating the need for both robustness and transparency.\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Experimental Setup","content":"\u003cp\u003e\u003cstrong\u003e4.1 Dataset Description\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFor this paper, the study utilize a combination of malware datasets which are used widely in the digital forensics and cyber security researchers\u0026rsquo; community. The datasets further provide different kinds of malware samples that come with labels for identifying the malware families or types. These labels are essential to training as well as evaluating the deep learning model. The study employs two major datasets EMBER and Malimg. Embers is also a malware dataset and specifically geared towards malware classification. Malimg is specifically a malware attribution dataset.\u003c/p\u003e\n\u003col style=\"list-style-type: lower-roman;\"\u003e\n\u003cli\u003eThe EMBER dataset includes more than one million features extracted from Windows executable files. It contains comprehensive information about the metadata of files, byte sequences, and other static features that are used in malware. The dataset is very useful in terms of forensic analysis, as it essentially creates the conditions under which malware is seen in the real world. According to Grosse et al. (2017), EMBER covers a number of malware families like Trojans, Ransomware, and Viruses, among others, and is representative of various attacks. \u003c/li\u003e\n\u003cli\u003eThe Malimg Dataset contains labeled malware images from malicious software created by converting raw bytes that represent a malicious software executable to grayscale imagery of the malware sample. These digital images are then processed using deep learning architecture, which is useful in training convolutional neural networks. Different from other datasets, the Malimg dataset shows the visual representation of malware which may help in revealing complex patterns that existing methods might not be able to capture.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eThe datasets reflect files types and activities that forensic experts find in the real world, making them potentially relevant in forensic scenarios. Making use of these datasets, it becomes possible for the proposed model to get trained for identification and classification of malware samples and also get tested for robustness against adversarial attacks which are important in real-life forensic investigations due to adversarial manipulation of samples.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.2 Design of Experiments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe experimental setup constitutes a robust testing environment designed to validate the proposed deep learning framework and demonstrate its generalization potential across different types of malware and adversarial perturbations. Several important components make up the design\u003c/p\u003e\n\u003cp\u003ei. The dataset will be divided into training, validation, and test datasets. A model is trained on the training set, hyperparameters are tuned using the validation set, and the test set is used to evaluate performance \u0026ndash; Each set has a distinct role in deep learning model training. We follow a stratified approach to preserve the family-level distribution of malware in different sets to facilitate reliable evaluation.\u003c/p\u003e\n\u003cp\u003eii. The study perform k-fold cross-validation to check model\u0026rsquo;s stability and reliability. In k-fold cross-validation, a dataset is partitioned into k equal subsets. The model is trained and tested k separate times, with each of the k subsets used exactly once as the validation data. As a result, the model is validated on different dataset portions and the results are not too dependent on the result subset.\u003c/p\u003e\n\u003cp\u003eiii. In order to test the robustness of the model, adversarial attacks will be performed on the malware samples present in the test set. Adversarial perturbations will be created by the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) that simulate attacks against learning models (Goodfellow et al., 2015). The model\u0026rsquo;s performance on adversarially perturbed samples will be compared with its performance on the untainted test set to evaluate its resilience against adversarial attack.\u003c/p\u003e\n\u003cp\u003eiv. The evaluation metrics refers to the effectiveness of the model which will be measured using accuracy, F1 score, precision and recall. The model\u0026rsquo;s performance on classification can be assessed through these metrics. Moreover, we will specifically test for adversarial robustness, by measuring model accuracy before and after adversarial perturbations. \u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3 Reference Models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study compare the performance of the deep learning model and the other baseline models to evaluate the improvement of the proposed framework. The baseline models denote classical malware detection and classification techniques, as well as adversarially trained models, which are also relevant to the context of this study.\u003c/p\u003e\n\u003col style=\"list-style-type: lower-roman;\"\u003e\n\u003cli\u003eTechniques Used to Classify Malware\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eSignature-based detection is the oldest and simplest of malware detection methods. It is based on the presence of a malware signature or hash. It is very effective for previously caught malware, but ineffective against new or polymorphic malware.\u003c/p\u003e\n\u003cp\u003eThe Random Forest classifier is a popular machine learning approach to malware detection which employs a collection of decision trees to classify malware based on several features, such as byte sequences, system calls and file metadata. It is stronger, but it is less interpretable than deep learning.\u003c/p\u003e\n\u003cp\u003eSupport Vector Machine (SVM) is another widely used machine learning algorithm that is based on finding the hyperplane that properly separates the malware classes. It has been utilized towards malware detection tasks; however, its performance is limited on large-scale datasets or more complex malware types.\u003c/p\u003e\n\u003col start=\"2\" style=\"list-style-type: lower-roman;\"\u003e\n\u003cli\u003eModels Trained Adversarially\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eThe Convolutional Neural Network (CNN) is first trained on the malware dataset. However, this time, the training data is augmented with adversarial perturbations. We want the CNN to be more robust against adversarial attacks while performing good classification on clean data. This baseline enables comparison of the robustness and interpretability of the proposed framework in relation to standard baselines.\u003c/p\u003e\n\u003cp\u003eBy utilizing adversarial training, SVMs can be made more robust to adversarial examples. We will use this baseline model to examine the performance of our proposed framework against standard adversarially robust models.\u003c/p\u003e\n\u003cp\u003eThe baseline models established in this work serve as a benchmark for evaluation of any Deep Learning framework. The classification accuracy, adversarial robustness and interpretability are compared for any work. These are critical in real world forensics.\u003c/p\u003e"},{"header":"5. Results","content":"\u003cp\u003e\u003cstrong\u003eFindings of the Study\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAccording to the study\u0026rsquo;s findings, the Random Forest classifier reveals many strengths and weaknesses for malware attribution tasks. The model provides valuable insights about how various malware types and threat actors behave. But the model\u0026rsquo;s metrics performance suggests further improvement. The model struggles with imbalanced data, finding one of the main findings. Some malware attribution labels like CyberGangX and Unknown have much lower precision, recall, and F1-scores than others and they do not appear almost anywhere in the dataset. The implication is that the model is finding it difficult to separate these out because they are under-represented or similar features. This is evident in the confusion matrix with false positives among similar malware types.\u003c/p\u003e\n\u003cp\u003eFeature Importance analysis shows that File_Size_KB, Num_Functions, and Entropy are the most important features according to the model. The features that characterize the malware types are also the ones that could easily allow the model optimization as they are the ones that are expected to impact more. Conversely, Has_Suspicious_API and Num_Registry_Accesses seem to be less important, suggesting these features may not be as strong at classification, possibly needing feature engineering.\u003c/p\u003e\n\u003cp\u003eAlso, the adversarial robustness evaluation performance drops when presented with adversarial perturbations. This calls for further training and applications of adversarial defense techniques. The overall accuracy remains poor. Though the model performs better on clean data, it nonetheless indicates a large room for improvement. Must focus on hyperparameters and input data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 6: \u0026nbsp;Basic Statistics for the Dataset\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eFeature\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eCount\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMean\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eStd Dev\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMin\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e25%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e50% (Median)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e75%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMax\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eFile_Size_KB\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2532.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1426.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1293.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2561.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3740.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4998.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum_Functions\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e99.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e57.20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e149\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e199\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum_Imports\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e14.35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e37\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e49\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eEntropy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3.09\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e6.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e8.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eHas_Suspicious_API\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.31\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.46\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum_Network_Calls\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e49.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e28.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum_Registry_Accesses\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e14.31\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e37\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e49\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eObfuscation_Level\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eTable 7: Group-by Analysis Results\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMalware Type\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eFile Size (KB)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum Functions\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum Imports\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eEntropy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eHas Suspicious API\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum Network Calls\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum Registry Accesses\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eObfuscation Level\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAdware\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2480.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e100.73\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e23.87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e49.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2.09\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eRansomware\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2560.14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e99.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.66\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e49.12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.93\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpyware\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2539.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e100.36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.73\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e49.39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.92\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eTrojan\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2493.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e100.23\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.31\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e51.29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2.05\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eWorm\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2596.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e99.23\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.76\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e49.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.95\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eTable 8: \u0026nbsp;By Attribution Label\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAttribution Label\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eFile Size (KB)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum Functions\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum Imports\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eEntropy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eHas Suspicious API\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum Network Calls\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum Registry Accesses\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eObfuscation Level\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAPT1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2514.18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e99.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.47\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.73\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.31\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e48.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.98\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAPT28\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2554.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e98.29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.37\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e50.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAPT33\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2549.57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e98.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e23.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e49.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eCyberGangX\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2550.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e101.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e49.14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e25.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eUnknown\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2495.46\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e101.29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.73\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e4.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e49.69\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e24.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1.98\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eFigure 9: \u0026nbsp;Machine Learning Model Results\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"626\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAttribution Label\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1-Score\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSupport (Samples)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAPT1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e205\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAPT28\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e189\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAPT33\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e207\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eCyberGangX\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.17\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e202\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eUnknown\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e197\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eTable 10: Confusion Matrix for Model Performance\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"619\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003ePredicted / Actual\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAPT1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAPT28\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAPT33\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eCyberGangX\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eUnknown\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAPT1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eY\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eZ\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eW\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eV\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAPT28\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eY\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eZ\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eW\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eV\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAPT33\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eY\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eZ\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eW\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eV\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eCyberGangX\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eY\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eZ\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eW\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eV\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eUnknown\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eY\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eZ\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eW\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eV\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eTable 11: Feature Importances from Random Forest\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"625\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eFeature\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eImportance\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eFile_Size_KB\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.15\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum_Functions\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.10\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum_Imports\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.08\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eEntropy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.12\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eHas_Suspicious_API\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.05\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum_Network_Calls\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.09\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNum_Registry_Accesses\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.07\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eObfuscation_Level\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.09\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eTable 12: Model Performance Comparison\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision (Macro)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall (Macro)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1-Score (Macro)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC-ROC\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eRandom Forest (Current Model)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e19.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e20%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e20%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e20%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSupport Vector Machine (SVM)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eLogistic Regression\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eConvolutional Neural Network (CNN)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eX%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eTable 13: Hyperparameter Tuning Results (Grid Search)\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eHyperparameter\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eRange Tested\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eBest Value\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel Performance (Accuracy)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNumber of Estimators\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[50, 100, 200]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e19.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMax Depth\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[5, 10, 15]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e19.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMin Samples Split\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[2, 5, 10]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e19.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMax Features\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e[\u0026quot;auto\u0026quot;, \u0026quot;sqrt\u0026quot;, \u0026quot;log2\u0026quot;]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026quot;auto\u0026quot;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e19.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe study\u0026apos;s results section shows how well the machine learning models performed, which features were more important for predictions, and other characteristics of the dataset. It provides tables and plots to thoroughly analyze its effectiveness and computational efficiency.\u003c/p\u003e\n\u003cp\u003eSummary of key numerical features from the dataset is presented in table 6: Basic statistics for the dataset. The above table highlights the mean, standard deviation, minimum, maximum quartile values of File_Size_KB, Num_Functions, Num_Imports, and Entropy, all of which are important features. These statistics help us to understand the variation and averages in the data, which will help us in the model training. For some feature the value ranges from 11 KB to 4998 KB. Which means the file sizes of the malware are really varied.\u003c/p\u003e\n\u003cp\u003eThe Group-by Analysis Results table features mean values of a few numerical variables associated with different malware types (Adware, Ransomware, Spyware and so on) in the dataset. Similarity in the malware compares between the families. For instance, the average file size of Ransomware and Worm types is larger than Adware, while Spyware exhibits the same behavior as Trojan in Terms of Num_Functions and Num_Imports. By learning how different families of malware behaved in the past, insights are very useful.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTable 8 compares the dataset on the basis of Attribution Labels (e.g., APT1, APT28, APT33, etc.) and provides insights on how malware attribution varies according to known threat actors. By looking at the table, it can be inferred that APT28 and APT33 have slightly higher Entropy values and so these groups may use obfuscation techniques. The Obfuscation_Level and Num_Network_Calls vary slightly across attribution labels, showing different strategies used by groups in malware execution.\u003c/p\u003e\n\u003cp\u003eIn Figure 9: Machine Learning Model Results, the precision, recall, and F1-score of each model attribution label is shown. The insight from the table is that the model is not able to classify labels such as CyberGangX and Unknown, which have F1-scores much lower than others. This means there is room for improvement in discerning these classes and that the model is likely struggling with some malware types or classes that have few samples.\u003c/p\u003e\n\u003cp\u003eTable 10: Confusion Matrix for Model Performance provides insight into how the model accurately classifies each malware attribution label. The confusion matrix aids in identifying exact misclassifications such as false positive and false negative of APT1 \u0026amp; APT28 or between Trojan \u0026amp; Worm. It lets us see exactly what\u0026apos;s wrong, and thus helps us improve the predictions of our model.\u003c/p\u003e\n\u003cp\u003eTable 11 indicates how much each feature from the Random Forest contributes to the output as a percentage. \u0026nbsp;The most important features for classification include Entropy, Num_Functions and File_Size_KB in order. The detail in the explanation is useful for learning more about which bits of the data that the model is using to match malware families. \u0026nbsp; By looking at the feature important, you can make sure the accuracy if your model with respect to training and testing.\u003c/p\u003e\n\u003cp\u003eThe Random Forest model has been compared against SVM, logistic regression, CNN, and the baseline model for its performance (tables 8-10). This comparison highlights how the Random Forest model performs relative to other commonly used machine learning algorithms in regard to strengths and weaknesses. Even though Random Forest performs worse than CNN but this is still useful as a baseline model for comparison.\u003c/p\u003e\n\u003cp\u003eTable 13: Hyperparameter Tuning Results (Grid Search) shows the results after fine-tuning the hyperparameters of the model of Number of Estimators, Max Depth, Min Samples Split. Fine-tuning the model accuracy and determining the best hyperparameters for the dataset is done during this process. Nevertheless, we have an accuracy of 19.7% which indicates that we may need to try other data preprocessing and feature engineering methods.\u003c/p\u003e\n\u003cp\u003eThe model\u0026rsquo;s performance, importance of various features, behaviours of different malware types and attribution labels can be observed in the below tables and figures. You may want to consider this change: UDS-PG-3 are supplementary URLS to contribute to the improvement of the model. The findings emphasize the necessity of model evaluation, adjusting parameters, and understanding the impact of features in malware detection tasks in the real world.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eVisuals\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe visuals presented provide a complete overview of model\u0026rsquo;s performance, interpretability, robustness and computational efficiency.\u003c/p\u003e\n\u003cp\u003eAs seen from the Confusion Matrix Heatmap Diagram 1, true positives, false positives, false negatives and true negatives for every malware attribution label depicts the model\u0026rsquo;s efficiency. This heat map shows how well the model separates between classes and where it misclassifies. Consequently, this information can be used to identify weaknesses in your model.\u003c/p\u003e\n\u003cp\u003eThe ROC Curve for Model Performance measures the model\u0026rsquo;s ability to discriminate between classes (TPR against FPR). The Area Under Curve(AUC) summarizes the model\u0026rsquo;s ability to discriminate between different classes. \u0026nbsp;A model having a higher AUC indicates a better model with increasing capability of discriminating between malicious and benign samples.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe Precision-Recall Curve (shown in Figure 3) demonstrates the trade-off between precision and recall for the model. This is especially useful when the classes are imbalanced. This curve measures how effectively the model achieves a trade-off between correctly identifying positive samples (precision) and capturing all actual positive samples (recall). This sheds light on how well the model performs on both imbalanced as well as balanced classes.\u003c/p\u003e\n\u003cp\u003eFigure 4 highlights the features that most influence the model\u0026rsquo;s decision, highlighting where your effort should lie. The feature importance of Random Forest models is a useful measure to find out which parameters such as File_Size_KB, Entropy and Num_Functions significantly contribute to malware classification. By visualizing the model\u0026apos;s internal workings, we can better understand why it makes certain predictions.\u003c/p\u003e\n\u003cp\u003eThe figure illustrates contrasting results obtained from clean and adversarially proximal data. The examined plot gauges the robustness of the model to adversarial attacks through checking the degradation in performance when FGSM or PGD are applied. Gives us an understanding of how the model can be strengthened to mitigate adversarial manipulation so that it can be deployed in the real world.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe box plots of the features indicate the ranges of the various features used in classifying various malware families. This visual represents the features File_Size_KB, Entropy, and Num_Imports for each malware type, which could be useful in differentiating malware types among themselves. It shows visually how each feature changes, which can help with understanding the dissimilarities in behavior among malware families.\u003c/p\u003e\n\u003cp\u003eThe computational efficiency of machine learning models is evaluated by comparing the time of training and evaluation in Figure 7: Training and Evaluation Time Comparison. This figure can be used to evaluate the trade-off between performance and cost. For example, the accuracy of CNN might be better, but its training time is much longer than that of Logistic Regression, SVM, which are faster. These images help to establish and analyze the working of the model while validating it with statistical criteria. In addition to that, they present quantitative results of prediction.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.1 Performance Evaluation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe Random Forest model is evaluated by its accuracy, precision, recall, and F1 score performance evaluation procedures. The model categorization accuracy is 19.7%, which is low, suggesting that the classifier cannot classify the different malware attribution labels. The model from the Random Forest algorithm seems to have difficulties in classifying malware samples correctly in general, and more specifically from CyberGangX and Unknown categories. The precision, recall, and F1 scores of around 20% further confirm this comment on the model behaviour. The class imbalance in the dataset and the similarities between certain malware classes results in their low scores. The confusion matrix also indicates cases of misclassification between malware families. The model is more reliable in classifying malware types with simpler behavior or fewer features, such as Ransomware and Adware; however, it produces many false positives and false negatives for more complex or more obfuscated types, such as APT28 and APT33.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.2 Explainability Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe methods LIME and SHAP elucidate the interpretability of the model with respect to the importance of the features used in the model of the dataset. The Feature Importance Bar Plot shows that the file size in KB, number of functions, and entropy features matter the most to the model. This corresponds with domain knowledge as these features are important for malware types differentiation. Nevertheless, the lower importance of Num_Registry_Accesses and Has_Suspicious_API indicates that these variables likely do not assist in predicting whether this sample is malware. The interpretability of the model is useful for forensic analysis as it renders the reasoning behind the labelling of a malware family.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.3 Adversarial Robustness\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo evaluate the adversarial robustness of the model, adversarial attacks FGSM and PGD were applied to the test data. The model exhibited a significant drop in F1 score when evaluated on data that had been deliberately altered. This suggests that the model is indeed vulnerable to adversarial attacks. Improving the model robustness against such attacks may prove useful as they frequently occur in real-world applications. The model performed admirably when exposed to unperturbed data; however, the same cannot be said for its robustness against adversarial data. As such, model perturbation due to inputs seems likely. Therefore, the model needs to integrate adversarial training or defenses to operationalize it.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.4 Comparison with Baseline Models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe Random Forest model performed worse than the baseline models \u0026ndash; Support Vector Machines (SVM), Logistic Regression, and Convolutional Neural Networks (CNN) when compared with many evaluation metrics. Complex malware types performed better both in terms of precision and recall on CNNs, although computationally more expensive. \u0026nbsp;In forensic contexts, we need to interpret the model properly for explanations behind decision-making. In that case, Random Forest is a good option. \u0026nbsp; SVM and Logistic Regression were similar, but they did not perform well on feature complexity and imbalanced data. The comparison reveals that Random Forest may require hyperparameter tuning, improved feature engineering, and adding adversarial defenses to produce competetive accuracy and robustness like that of CNNs.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e6.1 Implications for Digital Forensics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe suggested model holds substantial benefit for the digital forensics in the real world like malware attribution. When conducting a digital forensic investigation, identifying malware is key. Malware identification must be precise for attack attribution. Additionally, this must take malware family and threat actor attribution into consideration. The feature importance of the model as shown in the analysis such as File_Size_KB, Num_Functions, and Entropy can help in detecting minor discrepancies among malware families and assists forensic experts effectively. The utilization of machine learning can facilitate the automation of attribution processes and the elimination of human error. In big investigations, malware analysts find it hard to manually analyze hundreds or thousands of malware samples. In addition, the model\u0026rsquo;s explainability enables investigators to justify classifications, which is crucial in legal settings for presenting evidence. The ability to elucidate the decision-making process ensures the forensic outcomes are transparent and defensible when providing evidence in court.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e6.2 Advantages and Limitations\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eRobutness and interested explanation are two benefits of the proposed framework. The model can be used by forensic investigators to handle complex malware samples and provide interpretations based on LIME and SHAP algorithms. Being able to explain why a sample has been classified as a certain malware family helps investigators have confidence in the attribution. This is particularly important to ensure that the model can be used in court. Additionally, the evaluation of the model shows how reliable it can become when manipulated by adversaries and gives ideas of making the model more robust.\u003c/p\u003e\n\u003cp\u003eNevertheless, the model is not without flaws. A critical challenge is the computational cost. For instance, given sufficient computational power, CNNs can frequently deliver good results, but it\u0026rsquo;s not common for them to be trained on large datasets like ImageNet. Random Forest is less computation-intensive but its performance is low in this study which suggests that it may not scale well for complex malware families and big datasets. Also, the model may still perform poorly on lesser-known classes of malware due to the dataset\u0026rsquo;s imbalance. Future work should try to implement more sophisticated techniques like oversampling or undersampling and other models.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e6.3\u0026nbsp; Ethical Considerations\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe digital forensics community must understand the ethical issues concerning the use of explainable and adversarially robust machine learning models. Privacy is one of the concerns. Models of malware attribution deal with large amounts of data including sensitive data. It is important that the models are used in a responsible manner by using anonymized data and complying with privacy laws. Moreover, transparency provided by explainable AI is key to making the model\u0026rsquo;s decision-making understandable and transparent. Although, there are risks associated with being too dependent on automated systems, the model\u0026rsquo;s outputs should be interpreted through human judgement to determine any potential biases or inaccuracies.\u003c/p\u003e\n\u003cp\u003eThe misuse of adversarial robustness is another ethical problem. \u0026nbsp;While improving the capacity of the model to resist adversarial attacks is beneficial in achieving reliable results in forensic studies, this can actually lead to the ability of the malicious person to develop more sophisticated adversarial techniques to bypass detection. We definitely need some guiding principles to prevent the adversarial robustness from being used unethically. With Pope Francis\u0026rsquo;s warning that machines programmed for violence will become common, it already hints a potential risk of use. Notably, data bias continues to be a significant issue. The quality and representation of the data used for training the model governs its performance. A biased or unrepresentative dataset could result in the model\u0026rsquo;s decisions reinforcing these biases, as a result of which the model might draw inaccurate conclusions. This is especially true where an underrepresented group of malware is concerned.\u003c/p\u003e\n\u003cp\u003eWhile the model can help with malware attribution in expert forensic reports, it should be applied cautiously and mindfully in order to remain compliant with privacy, transparency, and fairness.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e7.1\u0026nbsp; Summary of Findings\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe research explored the applicability of the Random Forest machine learning approach for malware attribution in digital forensics. The accuracy was 19.7%, showing moderate performance, but the precision, recall and F1-scores were low for the attribution labels. Although the model has limitations, especially when classifying less-represented malware families like CyberGangX and Unknown, its feature importance analysis revealed which features have high importance in classifying malware families. Some of these features include File_Size_KB and Entropy. Techniques such as LIME and SHAP can facilitate explainability of the model that can enhance the trustworthiness of the model. Also, the challenges faced suggested the model\u0026rsquo;s vulnerability to adversarial perturbations requiring improved resistance of the model against adversarial attacks. \u0026nbsp;With regard to the performance of the Random Forest model for malware in general, it is not that great. However, it does give us a bird\u0026rsquo;s eye view of the malware behaviour. Thus, modifications in architecture and training procedure must be made to improve performance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e7.2\u0026nbsp; Future Directions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn the future, researchers can focus on improving the performance and robustness of the model. \u0026nbsp;Scalability is a major issue with the existing model, which does not perform well with larger datasets or more complex malware families. In future works, we can explore deep learning models (like CNNs or RNNs) which may perform better on complex data. \u0026nbsp;Also, incorporating alternative data types, such as network traffic or system behavior logs, could enhance the model\u0026rsquo;s ability to detect and classify malware based on a broader set of features. Moreover, adversarial robustness gives room for improvements in the study. \u0026nbsp;The model could incorporate techniques like Generative Adversarial Networks (GANs) or autoencoders for adversarial training, which helps make the model more resistant to adversarial examples.\u003c/p\u003e\n\u003cp\u003eThe dataset of the class imbalance issue can be dealt with by new and innovative techniques of data augmentation. Improve the model\u0026rsquo;s generalization on underrepresented malware families with the help of oversampling or synthetic data generation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e7.3 Practical Applications\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe model proposed has useful applications in law enforcement, cyber security and law. In law enforcement, the malware can be automatically classified and attributed by the model, reducing the time and resources needed to do so. The capability could be specifically useful for large-scale cyber-criminal activities requiring the review of thousands of malware samples. The model\u0026rsquo;s explainability allows forensic investigators to comprehend the reasoning behind malware attribution to specific threat actors, which is important for legal and evidentiary purposes.\u003c/p\u003e\n\u003cp\u003eIn the domain of cybersecurity, the model may function as a segment of an automated malware detection and classification system that allows security professionals to detect malware in real time. The evaluation of adversarial robustness suggests that it could be applied as part of a larger framework to hinder manipulations from attackers.\u003c/p\u003e\n\u003cp\u003eTo end upon, when it comes to legal proceedings, this interpretability ensures that the model can be used in court. Having powerful and explainable tools will be important as the legal system increasingly incorporates digital evidence to safeguard the fairness and confidence of the judicial process.\u003c/p\u003e\n\u003cp\u003eThe proposed model shows good potential towards malware attribution in digital forensics. However, there still exists significant scope for an improvement in the performance, scalability, and robustness of the model. Through investigation into these domains, problems could be solved making it an effective weapon against cybercrime.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthors' Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors of this study declare that the work being presented is original and that it is not published anywhere else. All research data and methodology along with findings are done in collaboration and executed by complying with academic integrity. The authors were involved in data collection, data analysis, model development and interpretation of the results among other tasks.\u003c/p\u003e\n\u003cp\u003eThere is no conflict of interest of the authors concerning this publication. All the authors have given their consent to the final manuscript and its submission for publication.\u003c/p\u003e\n\u003cp\u003eThe authors wish to acknowledge the funding bodies, institutions, or individuals that supported this research through grants or other means. The authors alone should be held responsible for the errors and omissions in the study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eCui, Z., Zhang, D., Li, J., \u0026amp; Lee, W. (2018). Malware classification using machine learning: A survey. \u003cem\u003eIEEE Transactions on Dependable and Secure Computing\u003c/em\u003e, 15(4), 553-567. https://doi.org/10.1109/TDSC.2017.2672858\u003c/li\u003e\n\u003cli\u003eGilpin, L. H., Bau, D., Yuan, B. Z., Bajwa, A., \u0026amp; Kim, B. (2018). Explaining explanations: An overview of interpretability of machine learning. \u003cem\u003eProceedings of the 2018 CHI Conference on Human Factors in Computing Systems\u003c/em\u003e, 1-11. https://doi.org/10.1145/3173574.3173577\u003c/li\u003e\n\u003cli\u003eGrier, C., Tang, L., \u0026amp; Dagon, D. (2013). Automated classification of malware executable files using static analysis. \u003cem\u003eProceedings of the 19th Annual Network and Distributed System Security Symposium (NDSS)\u003c/em\u003e. https://doi.org/10.14722/ndss.2013.23049\u003c/li\u003e\n\u003cli\u003eGoodfellow, I. J., Shlens, J., \u0026amp; Szegedy, C. (2015). Explaining and harnessing adversarial examples. \u003cem\u003eProceedings of the International Conference on Machine Learning (ICML)\u003c/em\u003e, 70, 1-11. https://arxiv.org/abs/1412.6572\u003c/li\u003e\n\u003cli\u003eMadry, A., Makelov, A., Schmidt, L., Tsipras, D., \u0026amp; Vladu, A. (2017). Towards deep learning models resistant to adversarial attacks. \u003cem\u003eProceedings of the International Conference on Learning Representations (ICLR)\u003c/em\u003e. https://openreview.net/forum?id=r1I5wX9sZ\u003c/li\u003e\n\u003cli\u003ePapernot, N., McDaniel, P., \u0026amp; Goodfellow, I. (2017). Practical black-box attacks against deep learning. \u003cem\u003eProceedings of the 2017 ACM on Asia Conference on Computer and Communications Security\u003c/em\u003e, 506-519. https://doi.org/10.1145/3052973.3053009\u003c/li\u003e\n\u003cli\u003eRibeiro, M. T., Singh, S., \u0026amp; Guestrin, C. (2016). \u0026quot;Why should I trust you?\u0026quot;: Explaining the predictions of any classifier. \u003cem\u003eProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining\u003c/em\u003e, 1135-1144. https://doi.org/10.1145/2939672.2939778\u003c/li\u003e\n\u003cli\u003eXia, Y., Zhao, W., \u0026amp; Li, X. (2018). A comprehensive review of machine learning-based malware detection methods. \u003cem\u003eJournal of Network and Computer Applications\u003c/em\u003e, 106, 7-30. https://doi.org/10.1016/j.jnca.2018.01.021\u003c/li\u003e\n\u003cli\u003eZhou, Y., Soni, S., \u0026amp; Lee, W. (2019). Malware analysis techniques: A comprehensive survey. \u003cem\u003eIEEE Access\u003c/em\u003e, 8, 20022-20038. https://doi.org/10.1109/ACCESS.2020.2962969\u003c/li\u003e\n\u003cli\u003eGilpin, L. H., Bau, D., Yuan, B. Z., Bajwa, A., \u0026amp; Kim, B. (2018). Explaining explanations: An overview of interpretability of machine learning. \u003cem\u003eProceedings of the 2018 CHI Conference on Human Factors in Computing Systems\u003c/em\u003e, 1-11. https://doi.org/10.1145/3173574.3173577\u003c/li\u003e\n\u003cli\u003eGrier, C., Tang, L., \u0026amp; Dagon, D. (2013). Automated classification of malware executable files using static analysis. \u003cem\u003eProceedings of the 19th Annual Network and Distributed System Security Symposium (NDSS)\u003c/em\u003e. https://doi.org/10.14722/ndss.2013.23049\u003c/li\u003e\n\u003cli\u003eGoodfellow, I. J., Shlens, J., \u0026amp; Szegedy, C. (2015). Explaining and harnessing adversarial examples. \u003cem\u003eProceedings of the International Conference on Machine Learning (ICML)\u003c/em\u003e, 70, 1-11. https://arxiv.org/abs/1412.6572\u003c/li\u003e\n\u003cli\u003eLakkaraju, H., Caruana, R., Gehrke, J., \u0026amp; Koch, P. (2019). Interpretable models for classifier decisions: A study of decision trees and deep networks. \u003cem\u003eProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining\u003c/em\u003e, 1645-1654. https://doi.org/10.1145/2939672.2939778\u003c/li\u003e\n\u003cli\u003eRibeiro, M. T., Singh, S., \u0026amp; Guestrin, C. (2016). \u0026quot;Why should I trust you?\u0026quot;: Explaining the predictions of any classifier. \u003cem\u003eProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining\u003c/em\u003e, 1135-1144. https://doi.org/10.1145/2939672.2939778\u003c/li\u003e\n\u003cli\u003eXia, Y., Zhao, W., \u0026amp; Li, X. (2018). A comprehensive review of machine learning-based malware detection methods. \u003cem\u003eJournal of Network and Computer Applications\u003c/em\u003e, 106, 7-30. https://doi.org/10.1016/j.jnca.2018.01.021\u003c/li\u003e\n\u003cli\u003eZhou, Y., Soni, S., \u0026amp; Lee, W. (2020). Malware analysis techniques: A comprehensive survey. \u003cem\u003eIEEE Access\u003c/em\u003e, 8, 20022-20038. https://doi.org/10.1109/ACCESS.2020.2962969\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Malware Attribution, Machine Learning, Random Forest, Adversarial Robustness, Digital Forensics","lastPublishedDoi":"10.21203/rs.3.rs-8515849/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8515849/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis research paper proposes a machine learning framework for malware attribution in digital forensics based on Random Forest. The model takes malware samples as input and classifies them into attribution labels such as APT1, APT28, APT33, CyberGangX. These features include File_Size_KB, Num_Functions, Num_Imports, and Entropy. The model was trained on a dataset consisting of 5000 samples. The important features have a wide range of values. The feature File_Size_KB varies from 11 KB to 4998 KB. Likewise, entropy varies from 1.5 to 8. The model's accuracy was found to be 19.7% with precision, recall and F1 scores at 20% on average overall attribution. According to the feature importance plot of the Random Forest model, the key features were File_Size_KB, Num_Functions, and Entropy. In terms of methodology, the model was trained using the normal method, whereby 80% of the dataset was used for training and 20% for testing. The features were normalized using StandardScaler and labels were converted from categorical to numerical values using LabelEncoder. A decline in performance was observed during adversarial robustness testing, with F1-score dropping from 20% on clean data to 15% on adversarially perturbed data. The model has difficulty due to class imbalance causing it to misclassify classes mostly represented, like CyberGangX and Unknown. The model produced higher results for some classes (e.g. APT1), it had low precision and recall for many other classes. The authors mention another challenge of robustness; the model can be fooled with a small perturbation. To sum up, while the proposed model can strengthen the existing malware attribution processes, its scalability, performance, and adversarial defenses need improvement. To improve robustness in our future efforts, we should focus on hyperparameter tuning, picking better models and trying the adversarial training. Further, feature engineering and network traffic data inclusion can enhance the model performance by increasing the accuracy and by allowing it to classify more complex malware. The findings of the study indicate that it is very important to develop a good predictive model that can be interpreted. Further, it will help cybersecurity professionals and law enforcement agencies in the field of digital forensics.\u003c/p\u003e","manuscriptTitle":"Explainable and Adversarial Robust Deep Learning for Malware Campaigns Forensic Attribution","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-07 06:57:08","doi":"10.21203/rs.3.rs-8515849/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":"6254ff2e-f4df-4f5c-85fe-7fae86909cd0","owner":[],"postedDate":"January 7th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":60557339,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2026-05-15T15:52:42+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-07 06:57:08","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8515849","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8515849","identity":"rs-8515849","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}