Selective adversarial learning for mobile malware

Machine learning models, including deep neural networks, have been shown to be vulnerable to adversarial attacks. Adversarial samples are crafted from legitimate inputs by carefully introducing small perturbation to the input so that the classifier is fooled. Adversarial retraining, which involves retraining the classifier using adversarial samples, has been shown to improve the robustness of the classifier against adversarial attacks. However, it has been also shown that retraining with too many samples can lead to performance degradation. Hence, a careful selection of the adversarial samples that are used to retrain the classifier is necessary, yet existing works select these samples in a randomized fashion. In our work, we propose two novel approaches for selecting adversarial samples: based on the distance from cluster center of malware and based on the probability derived from a kernel based learning (KBL). Our experiment results show that both of our selective mechanisms for adversarial retraining outperform the random selection technique and significantly improve the classifier performance against adversarial attacks. In particular, selection with KBL delivers above 6% improvement in detection accuracy compared to random selection. The method proposed here has greater impact in designing robust machine learning system for security applications.