Towards robust malware detection

Abstract

A central challenge of malware detection using machine learning methods is the presence of adversarial variants, small changes to detectable malware that allow it to evade a model (i.e. be classified as benign). We take inspiration from adversarial variant generation methods in the continuous-valued image domain to introduce methods for malware in the binary domain. We incorporate these methods in the training of hardened models towards the goal of robustness against adversarial variants. Additionally, we provide visualization tools for analysis of hardened models. Our tools illustrate the difference in loss behavior between models trained with different methods, the effect of adversarial learning on the loss landscape of a model, and the effect of adversarial learning on the decision map of a model. The adversarial learning framework and the visualization tools in combination allow for the creation and understanding of robust models.