Malware feature selection and adversarial sample generation method based on reinforcement learning
Xiaolong Li, Zhenhua Yan, Shuang Zhang, Xinghua Li, Feng Wang · 2024
Amidst the intricate battleground where attackers and defenders are perpetually locked in a sophisticated digital cat-and-mouse game, the task of discerning malware within an adversarial crafted environment manifests escalating complexity. The current research delineates the utilization of an enhanced reinforcement learning technique, orchestrated to engender adversarial malware specimens, thereby strategically navigating through machine learning detectors. This endeavor not only bolsters the robustness of malware identification systems but also adeptly navigates the perpetually evolving machinations of malware creators. Within this research, an environmental model is meticulously constructed, emulating detection engines and feature extractors, with malware samples assimilated as input. By integrating an autonomously generated reward function, we ascertain the model’s agility and the concomitant generation of manifold adversarial malevolent samples. The empirical evaluations underscore that, in contrast to conventional machine learning approaches, our methodology exudes superior flexibility and efficacy, furnishing a more formidable challenge to malware detection mechanisms.