Generating Adversarial Malware Using GANs to Evade Robust Detectors

Abstract

Malicious software is one of the most significant challenges in computer security. Continuous efforts to detect malicious software have evolved significantly since the advent of computing. With the rise of artificial intelligence, novel detection methodologies have emerged. This study investigates the application of Generative Adversarial Networks (GANs) to create new adversarial data capable of evading existing malware detectors. Our approach leverages GANs to modify the imported Dynamic-Link Libraries (DLLs) and Application Programming Interfaces (APIs) of malware files. Furthermore, our evaluation framework employs a detector trained on a comprehensive feature set including opcodes, PE-header attributes, and N-gram sequences-features that remain unmodified by the generator-to rigorously assess whether adversarial modifications to DLL/API features alone are sufficient to evade detection. The results demonstrate that even with a more robust detector, the GANs successfully generated adversarial attacks, achieving detection rates as low as 0%. This highlights the GANs network as a highly robust model capable of circumventing detectors, even when modifications are not directly applied to all features used by the detector. Moreover, these newly generated malware samples can serve as valuable training data for improving future malware detection systems, enabling them to better recognize and counter evolving threats.