Rethinking Lipschitzness Data-free Backdoor Defense

Zhiyu Zhu, Zhibo Jin, Huaming Chen, Teng Joon Lim

Keywords: backdoor attack, backdoor defense, model implement, system security, AI security

Abstract:

Deep Neural Networks (DNNs) have demonstrated remarkable success across various applications, yet some studies reveal their vulnerability to backdoor attacks, where attackers manipulate models under specific conditions using triggers. It significantly compromise the model integrity.
Addressing this critical security issue requires robust defence mechanisms to ensure the reliability of DNN models. However, most existing defence mechanisms heavily rely on specialized defence datasets, which are often difficult to obtain due to data privacy and security concerns. This highlights the urgent need for effective data-free defence strategies. In this work, we propose Lipschitzness Precise Pruning (LPP), a novel data-free backdoor defence algorithm that leverages the properties of Lipschitz function to detect and mitigate backdoor vulnerabilities by pruning neurons with strong backdoor correlations while fine-tuning unaffected neurons. Our approach optimizes the computation of the Lipschitz constant using dot product properties, allowing for efficient and precise identification of compromised neurons without the need of clean defence data. This method addresses the limitations of existing data-free defences and extends the scope of backdoor mitigation to include fully connected layers, ensuring comprehensive protection of DNN models. As our approach does not require data exchange, it can be implemented efficiently and effectively in diverse environments. Extensive experiments demonstrate that LPP outperforms state-of-the-art defence approaches without the need for additional defence datasets. We release our code at: https://anonymous.4open.science/r/LPP-CD3C.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9387