An Adaptive Defense Against Adversarial Patch Attacks For Vision Transformers
Keywords: vision transformer; adversarial patch attack; adptive defense
Abstract: Vision Transformers (ViTs) have become the prominent architecture for various computer vision tasks due to their superior ability to capture long-range dependencies through the self-attention mechanism. However, recent research indicates that ViTs are highly susceptible to carefully crafted adversarial patch attacks, presenting a significant challenge for practical deployment, particularly in security-critical applications. Existing approaches towards robust ViT frameworks often sacrifice clean accuracy and/or achieve suboptimal robustness, likely due to their uniform handling of diverse input samples. In this paper, we present NeighborViT, a novel adaptive defense framework specifically designed to counter adversarial patch attacks for ViTs. NeighborViT stands out by detecting and categorizing different types of attacks on inputs and applying adaptive, tailored defense mechanisms for each type of attack. To realize effective attack detection, categorization, and mitigation, NeighborViT explores the information in neighbor patches of the target patch and strategically employs them for defense. Our experimental results on the ImageNet dataset using various state-of-the-art ViT models demonstrate that NeighborViT significantly enhances robust accuracy without compromising clean accuracy. Our code is available at https://anonymous.4open.science/r/NeighborViT-8255.
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: 7095
Loading