Skip to content
InfoResearchPeer-reviewed

Pixel-Adaptive Purification: From Uniform to Spatially Heterogeneous Noise Injection

Published
Record updated
View JSON

Summary

Diffusion-based adversarial purification removes attacks by adding noise and then denoising, but uniform noise either over-corrupts clean regions or under-purifies perturbed ones. This research proposes a pixel-adaptive noise schedule guided by an SNR map, with an automatic stopping criterion and a two-stage latent optimization module. Experiments on CIFAR-10, CIFAR-100, SVHN, and ImageNet report that it outperforms representative adversarial training and purification methods under most attacks, including a stronger adaptive white-box setting, while keeping high standard accuracy.