Pulsar: Secure Steganography for Diffusion Models

Widespread efforts to subvert access to strong cryptography has renewed interest in steganography, the practice of embedding sensitive messages in mundane cover messages. Recent efforts at provably secure steganography have focused on text-based generative models and cannot support other types of models, such as diffusion models, which are used for high-quality image synthesis. In this work, we study securely embedding steganographic messages into the output of image diffusion models. We identify that the use of variance noise during image generation provides a suitable steganographic channel. We develop our construction, Pulsar, by building optimizations to make this channel practical for communication. Our implementation of Pulsar is capable of embedding ≈320--613 bytes (on average) into a single image without altering the distribution of the generated image, all in < 3 seconds of online time on a laptop. In addition, we discuss how the results of Pulsar can inform future research into diffusion models. Pulsar shows that diffusion models are a promising medium for steganography and censorship resistance.