CleanerCLIP: Fine-grained Counterfactual Semantic Augmentation for Backdoor Defense in Contrastive Learning

Yuan Xun, Siyuan Liang, Xiaojun Jia, Xinwei Liu, Xiaochun Cao

Keywords: backdoor defense, contrastive learning, multimodal pretrained models

Abstract:

Multimodal contrastive models like CLIP are increasingly vulnerable to data-poisoning backdoor attacks. Existing defense methods primarily target the pretraining phase. However, with the rise of open-source communities, pretrained models are now freely available for download and fine-tuning. These models may carry unknown security risks, posing significant threats to downstream users. This highlights the need for lightweight defense strategies tailored specifically for the fine-tuning stage. Current defenses during fine-tuning include: finetuning with clean data; and using unimodal self-supervised techniques like CleanCLIP, which has represented the state-of-the-art (SOTA). However, these methods rely on strengthening clean feature representations to mitigate attacks, making them ineffective against more stealthy backdoor techniques, such as BadCLIP, which leverage covert toxic features.
To overcome this limitation, we propose a finetuning defense mechanism based on fine-grained counterfactual text semantic augmentation. By modifying small portions of text during fine-tuning, our approach disrupts the association between backdoor triggers and target features.
We evaluate our method against six attack algorithms and conduct comprehensive zero-shot classification on ImageNet1K. Experimental results demonstrate that our method achieves SOTA performance in fine-tuning defense. Specifically, when facing the novel BadCLIP attack, our method surpasses CleanCLIP, reducing the Attack Success Rate (ASR) by 52.02% in the Top-1 and 63.88% in the Top-10 classifications.

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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 3105