Unlocking The Potential of Adaptive Attacks on Diffusion-Based Purification

Andre Kassis, Urs Hengartner, Yaoliang Yu

Abstract: Diffusion-based purification (DBP) has emerged as a cornerstone defense against adversarial examples (AEs), widely regarded as robust due to its use of diffusion models (DMs) that project AEs onto the natural data distribution. However, contrary to prior assumptions, we theoretically prove that adaptive gradient-based attacks nullify this foundational claim, effectively targeting the DM rather than the classifier and causing purified outputs to align with adversarial distributions. This surprising discovery prompts a reassessment of DBP's robustness, revealing it stems from critical flaws in backpropagation techniques used so far for attacking DBP. To address these gaps, we introduce DiffBreak, a novel and reliable gradient library for DBP, which exposes how adaptive attacks drastically degrade its robustness. In stricter majority-vote settings, where classifier decisions aggregate predictions over multiple purified inputs, DBP retains partial robustness to traditional norm-bounded AEs due to its stochasticity disrupting adversarial alignment. However, we propose a novel adaptation of a recent optimization method against deepfake watermarking, crafting systemic adversarial perturbations that defeat DBP even under these conditions, ultimately challenging its viability as a defense without improvements.