Evaluating a Malware Benchmark Dataset Using Neural Network Models and Shapley-Based Explainability

We present an analysis of the 2024 released MABEL dataset, a malware benchmark designed to evaluate machine learning models for malware detection and classification. We applied a three-layered neural network model to predict malware families based on the MABEL feature set, achieving an accuracy of $\mathbf{93 \%}$ for the original malware family names. We regrouped the malware families according to functionality, intent, and persistence mechanisms in order gain insights of malware behavior. We trained the same neural network model on these newly defined groups, achieving an accuracy of 81%. We applied Shapley-based explainability analysis to determine the contribution of each feature to the model’s predictions. Our findings reveal that the top features identified by Shapley values are highly consistent across different representations-whether predicting original family names or the regrouped categories. We conclude that the MABEL top features capture the essential characteristics of malware and provide robust insights into malware behavior.