MalwareTotal: Multi-Faceted and Sequence-Aware Bypass Tactics against Static Malware Detection

MalwareTotal: Multi-Faceted and Sequence-Aware Bypass Tactics against Static Malware Detection

연구 분야: Strategies

학회: ICSE '24: Proceedings of the IEEE/ACM 46th International Conference on Software Engineering

Recent methods have demonstrated that machine learning (ML) based static malware detection models are vulnerable to adversarial attacks. However, the generated malware often fails to generalize to production-level anti-malware software (AMS), as they usually involve multiple detection methods. This calls for universal solutions to the problem of malware variants generation. In this work, we demonstrate how the proposed method, MalwareTotal, has allowed malware variants to continue to abound in ML-based, signature-based, and hybrid anti-malware software. Given a malicious binary, we develop sequential bypass tactics that enable malicious behavior to be concealed within multi-faceted manipulations. Through 12 experiments on real-world malware, we demonstrate that an attacker can consistently bypass detection (98.67%, and 100% attack success rate against ML-based methods EMBER and MalConv, respectively; 95.33%, 92.63%, and 98.52% attack success rate against production-level anti-malware software ClamAV, AMS A, and AMS B, respectively) without modifying the malware functionality. We further demonstrate that our approach outperforms state-of-the-art adversarial malware generation techniques both in attack success rate and query consumption (the number of queries to the target model). Moreover, the samples generated by our method have demonstrated transferability in the real-world integrated malware detector, VirusTotal. In addition, we show that common mitigation such as adversarial training on known attacks cannot effectively defend against the proposed attack. Finally, we investigate the value of the generated adversarial examples as a means of hardening victim models through an adversarial training procedure, and demonstrate that the accuracy of the retrained model against generated adversarial examples increases by 88.51 percentage points.