Vision Transformers (ViTs) achieve superior performance on various tasks compared to convolutional neural networks (CNNs), but ViTs are also vulnerable to adversarial attacks. Adversarial training is one of the most successful methods to build robust CNN models. Thus, recent works explored new methodologies for adversarial training of ViTs based on the differences between ViTs and CNNs, such as better training strategies, preventing attention from focusing on a single block, or discarding low-attention embeddings. However, these methods still follow the design of traditional supervised adversarial training, limiting the potential of adversarial training on ViTs. This paper proposes a novel defense method, MIMIR, which aims to build a different adversarial training methodology by utilizing Masked Image Modeling at pre-training. We create an autoencoder that accepts adversarial examples as input but takes the clean examples as the modeling target. Then, we create a mutual information (MI) penalty following the idea of the Information Bottleneck. Among the two information source inputs and corresponding adversarial perturbation, the perturbation information is eliminated due to the constraint of the modeling target. Next, we provide a theoretical analysis of MIMIR using the bounds of the MI penalty. We also design two adaptive attacks when the adversary is aware of the MIMIR defense and show that MIMIR still performs well. The experimental results show that MIMIR improves (natural and adversarial) accuracy on average by 4.19\% on CIFAR-10 and 5.52\% on ImageNet-1K, compared to baselines. On Tiny-ImageNet, we obtained improved natural accuracy of 2.99\% on average and comparable adversarial accuracy. Our code and trained models are publicly available\footnote{\url{https://anonymous.4open.science/r/MIMIR-5444/README.md}}.
翻译:暂无翻译