Diff-Cleanse: Identifying and Mitigating Backdoor Attacks in Diffusion Models

Diffusion models (DMs) are advanced generative models, yet recent research reveals their vulnerability to backdoor attacks, which establish hidden associations between input patterns and targeted model behavior, potentially causing malicious outputs during inference. These attacks pose significant risks, including model owner reputation damage and harmful content generation. However, existing defense methods often fail against backdoor attacks on diffusion models. To address this gap, we propose Diff-Cleanse, a two-stage defense framework. The first stage introduces a novel trigger inversion method for backdoor detection, and the second stage applies a structural pruning-based method for backdoor removal. Experiments on 373 models poisoned by three state-of-the-art attacks show that Diff-Cleanse achieves > 97% detection accuracy, completely remove backdoors and maintains the models' benign performance. Code is available at https://github.com/shymuel/diff-cleanse.