Universal adversarial purification for Stable Diffusion based on DDIM metric loss (UDAP)

The researchers propose the UDAP framework, which removes adversarial noise from Stable Diffusion training data by minimizing a DDIM metric loss. It exploits the difference in reconstruction behavior between clean and adversarial images in DDIM inversion, and uses a dynamic epoch adjustment strategy to iterate adaptively according to reconstruction error, balancing efficiency and purification quality. Experiments show that UDAP can resist attacks including PID targeting the VAE, Anti-DreamBooth targeting the UNet, the hybrid attack MIST, Anti-Diffusion, and MetaCloak, and generalizes across SD versions and prompts.