Watch Your Step: Information Injection in Diffusion Models via Shadow Timestep Embedding

Researchers from arXiv (2605.00935) introduce Shadow Timestep Embedding (STE), a covert information injection attack targeting diffusion models by exploiting the underutilized temporal conditioning space of timestep embeddings. STE encodes side-channel data into distinct per-timestep representational capacities without degrading generation quality, creating a steganographic backdoor invisible to standard audits. Security engineers and model provenance researchers should treat this as a critical attack surface, particularly for open-weight diffusion deployments. Unlike prior backdoor methods targeting weights or training data, STE operates purely within the temporal embedding pipeline, making detection significantly harder.