Interpretable Machine Learning for Mitigating Feature-Driven Attacks

• Published in: IEEE transactions on technology and society Vol. 6; no. 2; pp. 220 - 230
• Main Authors: Hartman, Corey M., Rimal, Bhaskar P.
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.06.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Analytical models; Bias; Codes; Context modeling; Data models; Datasets; Documents; Explainable AI; Feature extraction; Heuristic algorithms; interpretable machine learning; Machine learning; macros; malicious documents; Malware; Mathematical models; Predictions; Predictive models; Robot learning; SHAP; VBA; XAI; XGBoost
• ISSN: 2637-6415; 2637-6415
• DOI: 10.1109/TTS.2025.3531780

Recent studies have found that 43% of malware infections begin as malicious Microsoft Office documents in the form of Word or Excel file. While many techniques are proposed and are effective in the detection of malicious documents through the utilization of machine learning (ML) algorithms, bias in the datasets and the lack of insight into the decision as to why a document was flagged as malicious are problematic, as one key feature focused on by the ML model utilized may be relied on solely for the prediction that is made. By utilizing the SHAP algorithm (SHapley Additive exPlanation) and an ensemble of ML algorithms split into groups by their SHAP magnitude, where those features taking over the decision-making process of a model are split into their own feature set and are utilized in the training of a separate ML model, a voting classifier can be made to reduce this bias and reliance on a single or select few features. That allows for a more robust ML model for predicting malicious Office documents and presenting more insight into why a prediction was made by the classifier and a model that can let the user know when not enough data is present to predict with confidence. By utilizing this technique, an ensemble soft voting classifier was created that obtained 90.1% accuracy on a balanced dataset consisting of 250 malicious and 250 benign randomly selected Office documents and presents the user with a simple natural language statement that indicates the classification of the documents and why it was classified as a specific label.