An Adversarial Attack on Artificial Intelligence Malware Detection in Consumer Internet of Things
Chi-Hsin Yang, Maina Bernard Mwangi, Shin‐Ming Cheng, Hahn-Ming Lee · IEEE Consumer Electronics Magazine · 2024
The proliferation of the Internet of Things (IoT) and artificial intelligence (AI) has transformed traditional consumer electronics (CEs) into next-generation devices with enhanced intelligence and connectivity. However, this advancement has exposed CEs to cybersecurity threats, such as IoT botnets. Consequently, researchers are employing AI for proactive threat detection and prevention. Unfortunately, AI algorithms are vulnerable to adversarial attacks, necessitating robustness studies, as evaded malware can cause significant damage to already susceptible IoT CEs. This article presents a case study to evaluate the resilience of AI-based IoT malware detection systems against adversarial attacks. Specifically, our method involves inserting crafted binary code snippets (payloads) into the empty regions of malware executables. We leverage explainable AI techniques to guide payload generation, coupled with an optimization procedure to efficiently identify optimal payload sequences. Our method, tested on real-world IoT datasets, yields a robust hybrid detection system with a detection rate of up to $99.11\%.$ Our attack approach achieves evasion rates of up to 100% and generates transferable adversarial examples. The generated samples evade a prominent structural IoT malware detector with an evasion rate of 95.15% at a minimal attack cost. This study underscores the importance of enhancing the robustness of AI-based malware detection systems and implementing diverse strategies to safeguard consumer IoT devices.