Evading PDF malware classifiers with generative adversarial network

Generative adversarial networks (GANs) have become one of the most popular research topics in deep learning. It is widely used in the term of image, and through the constant competition between generator and discriminator, it can generate so remarkably realistic images that human can’t distinguish. However, Although GAN has achieved great success in generating images, it is still in its infancy in generating adversarial malware examples. In this paper, we propose an PDF malware evasion method that is using GAN to generate adversarial PDF malware examples and evaluate it against four local machine learning based PDF malware classifiers. The evaluation is conducted on the same dataset which contains 100 malicious PDF files. The experimental results reveal that the proposed evasion attacks are effective, with attacks against three classifiers all attaining 100% evasion rate and attack against the last classifier also attaining 95% evasion rate on the evaluation dataset.