Abstract
Adversarial patch attacks mislead neural networks by injecting adversarial pixels within a designated local region. Patch attacks can be highly effective in a variety of tasks and physically realizable via attachment (e.g. a sticker) to the real-world objects. Despite the diversity in attack patterns, adversarial patches tend to be highly textured and different in appearance from natural images. We exploit this property and present PatchZero, a task-agnostic defense against white-box adversarial patches. Specifically, our defense detects the adversarial pixels and "zeros out" the patch region by repainting with mean pixel values. We formulate the patch detection problem as a semantic segmentation task such that our model can generalize to patches of any size and shape. We further design a two-stage adversarial training scheme to defend against the stronger adaptive attacks. We thoroughly evaluate PatchZero on the image classification (ImageNet, RESISC45), object detection (PASCAL VOC), and video classification (UCF101) datasets. Our method achieves SOTA robust accuracy without any degradation in the benign performance.
Abstract (translated)
URL
https://arxiv.org/abs/2207.01795