TEMPO-Diffusion: Temporally Exposed Malicious Poisoning of Diffusion Models
TEMPO-Diffusion introduces a supply-chain poisoning attack targeting diffusion models used to generate synthetic training data for downstream classifiers. By embedding malicious triggers within the temporal denoising process rather than the spatial input, the attack ensures synthetic outputs remain visually indistinguishable from clean data, maintaining low Fréchet Inception Distance (FID) scores. This allows adversaries to inject stealthy backdoors into downstream computer vision models, such as those used in autonomous driving (CALISA, GTSRB), enabling targeted misclassifications. The vulnerability exploits the reliance of MLOps pipelines on foundation models, localizing the distribution shift within reverse diffusion timestamps to bypass standard visual inspection and data sanitization.