Adversarial-Example Attacks Toward Android Malware Detection System

Recently, it was shown that the generative adversarial network (GAN) based adversarial-example attacks could thoroughly defeat the existing Android malware detection systems. However, they can be easily defended through deploying a firewall (i.e., adversarial example detector) to filter adversarial examples. To evade both malware detection and adversarial example detection, we develop a new adversarial-example attack method based on our proposed bi-objective GAN. Experiments show that over $\text{95\%}$ of adversarial examples generated by our method break through the firewall-equipped Android malware detection system, outperforming the state-of-the-art method by $\text{247.68\%}$ .