We propose a novel adversarial purification technique based on optimization that regards the attack landscape.

DiffBreak provides the first reliable framework for differentiating through diffusion-based purification, revealing key vulnerabilities under adaptive attacks.