Adversarial Attacks on Mobile Malware Detection

In recent years, machine learning approaches have been widely adopted for computer security tasks, including malware detection. Malware is a potent threat and an ongoing issue especially on smartphones which account for more than half of global web traffic. Although detection solutions are improving with the advances in machine learning techniques, they have been shown to be vulnerable to adversarial samples that carefully crafted perturbation enables them to evade detection. We propose a machine learning based model to attack malware classifiers leveraging the expressive capability of generative adversarial networks (GANs). We use GANs to generate effective adversarial samples by implying a threshold on the distortion amount on the generated samples. We show that the generated samples can bypass detection in 99% of attempts using a real Android application dataset.