Researchers unveil InvCISD: AI steganography with diffusion models

A new paper introduces InvCISD, an invertible diffusion framework for coverless image steganography (CIS), which synthesizes stego images rather than modifying existing cover images so authorized recipients can reconstruct the original secret image. Existing diffusion-based CIS methods produce natural-looking stego images but retain strong visual similarity to the secret, potentially exposing structural and semantic cues. InvCISD addresses this by coupling the latent representations of the secret and an irrelevant reference image using an invertible network called LIMNet, first training LIMNet in diffusion latent space, then fine-tuning the full network end-to-end. Experiments show the method substantially reduces secret-stego visual similarity, improves stego quality, and maintains satisfactory secret reconstruction quality. The article also reports that all evaluated methods were highly detectable by a CIS-oriented steganalysis model, suggesting that resisting targeted steganalysis remains a critical direction for future research.