Robustness of Image-Based Malware Classification Models trained with Generative Adversarial Networks

Abstract

As malware continues to evolve, deep learning models are increasingly used for malware detection and classification, including image-based classification. However, adversarial attacks can be used to perturb images so as to evade detection by these models. This study investigates the effectiveness of training deep learning models with Generative Adversarial Network-generated data to improve their robustness against such attacks. Two image conversion methods, byteplot and space-filling curves, were used to represent the malware samples, and a ResNet-50 architecture was used to train models on the image datasets. The models were then tested against a projected gradient descent attack. It was found that without GAN-generated data, the models’ prediction performance drastically decreased from 93-95% to 4.5% accuracy. However, the addition of adversarial images to the training data almost doubled the accuracy of the models. This study highlights the potential benefits of incorporating GAN-generated data in the training of deep learning models to improve their robustness against adversarial attacks.