ConsistencyPure: Consistency Models for Realtime Adversarial Purification

Diffusion-Based Purification (DBP) is a stochasticity-based defense method that effectively mitigates adversarial attacks but suffers from high computational costs due to its iterative reverse process. To address this, we propose ConsistencyPure, a novel purification framework leveraging a pre-trained consistency model distilled from diffusion models. ConsistencyPure requires only one forward and one reverse sampling step, achieving real-time purification while maintaining comparable defense performance and improving efficiency by one to two orders of magnitude. Our finding further confirms that the robustness of DBP stems primarily from randomness, independent of the diffusion framework’s iterative operations or noise prediction. Importantly, we highlight for the first time the crucial role of model representation learning in stochasticity- based defense methods. Extensive experiments validate the effectiveness and practicality of ConsistencyPure.