extraction-attack

Cyclic denoising is introduced as a novel extraction attack that reveals ultrastable memorized training images in diffusion models by repeatedly noising and denoising samples. The technique requires no gradients or weight inspection and has implications for privacy auditing.