UIBDiffusion: Universal Imperceptible Backdoor Attack for Diffusion Models
Published in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025
Applied a specific imperceptible trigger, built from universal adversarial perturbations, to part of the images in a training dataset to poison it. The poisoned diffusion model exhibits significant, controllable performance deviations when generating target images upon detecting the trigger, while keeping generation quality high on clean data. The trigger is image- and model-agnostic, and is shown to bypass state-of-the-art backdoor defenses for diffusion models (Elijah, TERD). Experiments were carried out on CIFAR-10 and CelebA-HQ 256, evaluated with FID, MSE, and SSIM.
Recommended citation: Han, Y., Zhao, B., Chu, R., Luo, F., Sikdar, B., & Lao, Y. (2025). "UIBDiffusion: Universal Imperceptible Backdoor Attack for Diffusion Models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).
Download Paper