PSyDUCK: hiding information in the denoising process of latent diffusion models

Recent advances demonstrate that information can be covertly embedded in the outputs of stochastic generative AI models, raising both opportunities for secure communication and risks of misuse. Existing latent diffusion steganography methods typically hide data in the entropy of the initial latent state, inherently limiting embedding capacity. In this work, we instead investigate information hiding within the entropy of the diffusion denoising process itself. We introduce PSyDUCK, a simple but efficient framework that leverages controlled divergence and local mixing during denoising to enable high-capacity message embedding while preserving visual fidelity. Our empirical evaluation shows PSyDUCK can hide substantial information in both image and video diffusion models. While our formal analysis indicates that the security guarantees of denoising-based embedding are limited, the existence of this channel nonetheless requires that steganalysis methods account for entropy throughout the entire denoising process - not just in the initial latent state.