In this paper, researchers focus on developing a secure steganography technique called Pulsar for diffusion models, commonly used in high-quality image synthesis. They highlight the limitations of text-based generative models in supporting steganography and propose using variance noise during image generation as a steganographic channel. The Pulsar construction allows for embedding approximately 320-613 bytes of data into a single image without altering its distribution, all in less than 3 seconds of online time on a laptop. The authors stress the potential of diffusion models for steganography and censorship resistance, providing valuable insights for future research in this area.
https://eprint.iacr.org/2023/1758