ViP: Unified Certified Detection and Recovery for Patch Attack with Vision Transformers

Patch attack, which introduces a perceptible but localized change to the input image, has gained significant momentum in recent years. In this paper, we present a unified framework to analyze certified patch defense tasks, including both certified detection and certified recovery, leveraging the recently emerged Vision Transformers (ViTs). In addition to the existing patch defense setting where only one patch is considered, we provide the very first study on developing certified detection against the dual patch attack, in which the attacker is allowed to adversarially manipulate pixels in two different regions. By building upon the latest progress in self-supervised ViTs with masked image modeling (i.e., masked autoencoder (MAE)), our method achieves state-of-the-art performance in both certified detection and certified recovery of adversarial patches. Regarding certified detection, we improve the performance by up to similar to 16% on ImageNet without training on a single adversarial patch, and for the first time, can also tackle the more challenging dual patch setting. Our method largely closes the gap between detection-based certified robustness and clean image accuracy. Regarding certified recovery, our approach improves certified accuracy by similar to 2% on ImageNet across all attack sizes, attaining the new state-of-the-art performance.