Patch Vestiges in the Adversarial Examples Against Vision Transformer Can Be Leveraged for Adversarial DetectionDownload PDF

22 Nov 2021, 06:35 (modified: 10 Dec 2021, 11:33)AAAI-22 AdvML Workshop ShortPaperReaders: Everyone
Keywords: adversarial examples, vision transformer, adversarial detection, patch vestiges
TL;DR: The division of the patches by ViT remains large enough vestiges in the adversarial examples to be a method of detection.
Abstract: Vision Transformer (ViT), a Transformer-based architecture that divides images into patches, can catch up with or surpass convolution-based networks in multiple Computer Vision tasks. However, ViT is also vulnerable in the face of adversarial examples (AEs). Thus the topic around the attack and defense of ViT becomes very rewarding. Recent studies have found that the AEs against ViT seem to have grid-like textures that coincide with the patches. In this paper we confirm such sensation is true. We show that these grid-like textures are the remained vestiges due to the patch division from ViT. We name them as Patch Vestiges. We propose statistics to measure the sizes of Patch Vestiges in the images or AEs quantitatively. We also build a linear regression classifier to detect the AEs against ViT practically via the proposed statistics. The experiments show that the performance of the simple classifier can even match some recent adversarial detection methods, suggesting that when trying to attack ViT or detect the AEs against ViT, Patch Vestiges are worth considering about as a critical factor.
2 Replies