Researchers introduce Shadow Timestep Embedding mechanism revealing timestep embeddings as a side channel for information injection in diffusion models

Researchers present Shadow Timestep Embedding (STE), a mechanism that investigates timestep embeddings—a component that provides temporal conditioning to the denoising network in diffusion models—for potential malicious information injection and side-channel attacks.

The work finds that different timesteps in diffusion models exhibit distinct representational capabilities that can encode side-channel information, and provides theoretical analysis of timestep embeddings as position-encoding mappings with a mutual coherence evaluation explaining separability of disjoint timestep intervals.

The findings position the diffusion model's timestep as a powerful side channel for carrying dedicated information, motivating new research directions for adversarial generative modeling by understanding the temporal dimension of these systems.