Beyond Pretrained Features: Noisy Image Modeling Provides Adversarial Defense
share
Masked Image Modeling (MIM) has been a prevailing framework for self-supervised visual representation learning. Within the pretraining-finetuning paradigm, the MIM framework trains an encoder by reconstructing masked image patches with the help of a decoder which would be abandoned when the encoder is used for finetuning. Despite its state-of-the-art performance on clean images, MIM models are vulnerable to adversarial attacks, limiting its real-world application, and few studies have focused on this issue. In this paper, we have discovered that noisy image modeling (NIM), a variant of MIM that uses denoising as the pre-text task, provides not only good pretrained visual features, but also effective adversarial defense for downstream models. To achieve a better accuracy-robustness trade-off, we further propose to sample the hyperparameter that controls the reconstruction difficulty from random distributions instead of setting it globally, and fine-tune downstream networks with denoised images. Experimental results demonstrate that our pre-trained denoising autoencoders are effective against different white-box, gray-box, and black-box attacks without being trained with adversarial images, while not harming the clean accuracy of fine-tuned models. Source code and models will be made available.
READ FULL TEXT