Active Learning Based Adversary Evasion Attacks Defense for Malwares in the Internet of Things

Abstract

In this article, we study adversarial evasion attacks in the context of an active learning environment. To prevent evasion attacks in Internet of Things environments, a feature subset selection method is proposed. To train an independent classification model for a single Android application, the approach extracts application-specific data from that application. We compare and evaluate the performance of Android malware benchmarks using ensemble-based active learning, followed by the use of a collaborative machine learning classifier to protect against adversarial evasion attacks on a dataset of Android malware benchmarks. It was found that the proposed approach generates 0.91 receiver operating characteristic with 14 fabricated input features.