Offense as Defense: Efficient Detection of Backdoor Attacks With Adversarial Samples in Cyber-Physical Social Systems

Hengqi Guo, Shijing Hu, Zhihui Lu, Zirui Guo, Haoran Zhou, Qiang DuanIEEE Transactions on Computational Social Systems 2025 · OpenReview

Abstract

Collaborative learning approaches in cyber-physical social systems (CPSSs) improve cooperation among entities; however, their framework for third-party collaboration is inherently vulnerable to backdoor attacks. These attacks involve the insertion of hidden triggers into the model for malicious purposes during the training process. A variety of efforts have been made to detect backdoors to ensure model security. However, current approaches typically leverage trigger reverse generation and anomalous neuron analysis and are not effective and efficient in detecting various backdoor attacks. To tackle this issue, we present an offense-as-defense (OAD) framework that creatively employs adversarial sample attacks as an approach to backdoor detection. The OAD framework is based on our observation that the existence of a backdoor in a model may change the model behavior under an adversarial sample attack, either strengthening or weakening the robustness of the model against the adversarial sample attack. Therefore, by quantitatively evaluating the magnitudes of change in model performance under adversarial sample attacks, the OAD framework is able to not only detect backdoors in models but also identify the specific backdoor attack methods. The effectiveness of the proposed OAD framework in defending against existing backdoor attacks has been demonstrated through extensive experiments.
BibTeX
@article{guo2025offense,
  author = {Hengqi Guo and Shijing Hu and Zhihui Lu and Zirui Guo and Haoran Zhou and Qiang Duan},
  title = {{Offense as Defense: Efficient Detection of Backdoor Attacks With Adversarial Samples in Cyber-Physical Social Systems}},
  journal = {IEEE Transactions on Computational Social Systems},
  year = {2025},
  url = {https://openreview.net/forum?id=wxFtViCfiM},
}

Loading related content...

Figure