Enhance the Transferability of Adversarial Attacks through Channel Pruning

Chunghao Liao, Shang-Tse Chen

Keywords: adversarial attacks transferability, channel pruning, model augmentation

Abstract: Recent studies have shown that neural networks are vulnerable to adversarial attacks, where attackers generate adversarial samples by imposing tiny noise. The tiny noise can not misguide human perception, though leading the neural networks to generate wrong predictions. Transfer-based black-box attacks play a more significant role in recent studies due to their more realistic setting and considerable progress in performance. Previous studies have shown that some different channels of the same layer in convolution neural networks (CNN) contain lots of repetitive information, and we find that existing transferable attacks tend to exploit those redundant features more, which limits their transferability. Hence, we advocate using channel pruning and knowledge distillation to conduct model augmentation. In addition, we introduce a method of regularization on the gradients of intermediate feature maps of augmented models, which further enhances the transferability of our method. Comprehensive experiments demonstrate that imposing our method of model augmentation on existing methods can significantly improve the transferability of adversarial attacks in untargeted or targeted scenarios. Furthermore, our method outperforms state-of-the-art model augmentation techniques without the usage of additional training datasets.

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: 4559