Generative Ensemble Learning for Mitigating Adversarial Malware Detection in IoT

Abstract

This paper proposes a framework that can be employed to mitigate adversarial evasion attacks on Android malware classifiers. It extracts multiple discriminating feature subsets from a single Android app such that each subset has the potential to classify a huge dataset of malicious and benign Android apps independently. Moreover, it incorporates an ensemble of ML classifiers where each classifier is trained on different features subset. Finally, the ensemble model formulates a collaborative classification decision that is resilient against adversarial evasion attacks. Results showed that the designed model achieves good performance compared to the existing models.