GAN-Based Adversarial Attack Against Malware Detection

Traditional malware detection methods based on signatures or behaviors are increasingly struggling to cope with new types of attacks, prompting a shift towards detection methods based on machine learning and deep learning. Compared to traditional detection mechanisms, machine learning and deep learning-based methods possess stronger feature extraction and generalization capabilities. This paper presents a malware detection model implemented using a Residual Neural Network (ResNet), capable of accurately distinguishing between malicious and benign software. However, detection systems based on machine learning or deep learning have issues with robustness, as malware authors can adjust the features of their malware to evade detection. We implemented an attack method based on Generative Adversarial Networks (GAN), targeting a trained ResNet detector, and used GAN to generate adversarial malware samples. The results demonstrate that this attack strategy can dynamically generate adversarial malware samples while realistically replicating actual attack scenarios.