Abstract:

In recent years, using machine learning algorithm to detect mobile terminal malware has become a research hotspot. In order to make the malware evade detection, malware producers use various methods to make malicious adversarial samples. This paper proposes an algorithm MalWGAN based on Wasserstein GAN (WGAN) to generate mobile terminal malicious adversarial samples, which can bypass the black box model detector based on machine learning algorithms to evade detection. Different from the existing adversarial samples generated by static gradient methods, the MalWGAN model combines API calls and static features to generate adversarial samples. Since adversarial samples are dynamically generated by the feedback of the black box model detector, the probability of escaping from the detection of the black box model detector is higher.

Key words: adversarial sample, WGAN, evasion of detection