Robustness of Image-based Android Malware Detection Under Adversarial Attacks
Asim Darwaish, Farid Naït‐Abdesselam, Chafiq Titouna, Sumera Sattar · 2021
The exhilarating pace of smartphone innovation and vast proliferation to everyday life also impose serious security threats. The open-source and largest android market is also the hive for malware authors. Since the last decade, machine learning (ML) has gained much attraction and successfully deployed and offers unsurpassed versatility for automated malware detection. Unfortunately, as ML-based approaches become widely adopted and deployed, adversaries are also in a never ending race to evade these classifiers for bypassing android malware detection systems. To combat adversarial attacks and secure machine learning-based android malware classifiers, we present a novel image based android malware classifier that has proven its robustness under various adversarial settings. In this work, we have crafted two novel attacks and reveal that the state of the art Android malware detectors are vulnerable and got easily evaded with more than 50% evasion rate. However, the proposed approach establishes a durable defense line against these adversarial attacks and too arduous to bypass.