AdVul: Adversarial Attack against ML-based Vulnerability Detection

A fundamental problem in software security, detecting software vulnerabilities (i.e. vulnerabilities), has achieved tremendous progress by leveraging machine learning (ML) techniques. These ML-based approaches provide an efficient solution for identifying vulnerabilities and have demonstrated high detection accuracy in recent studies. However, it remains unclear to what extent ML-based vulnerability detection is vulnerable to widespread adversarial attacks. In this paper, we propose a framework that performs Adversarial Attack against ML-based Vulnerability Detection (AdVul) to investigate the vulnerability of the ML-based vulnerability detection. Our experiments on three representative ML models for vulnerability detection demonstrate that the proposed AdVul attack can fool victim models with a high success rate while maintaining the functionality and performance of original programs. Furthermore, we discuss two potential defense methods, enhanced data preprocessing and adversarial training. Our evaluation demonstrates that both methods can to a meaningful extent mitigate the proposed AdVul adversarial attacks. More broadly, our work intends to prompt the security community to seriously consider the potential risks associated with the extensive use of ML techniques in the security domain.