InvCISD: Invertible Diffusion Framework for Coverless Image Steganography

A recent study published on arXiv (ID: 2608.13597) presents InvCISD, a novel invertible diffusion framework designed for coverless image steganography (CIS). Unlike traditional methods that alter an existing cover image, CIS creates a stego image, enabling authorized users to retrieve the original secret image. While current diffusion-based CIS techniques produce visually appealing stego images, they often maintain significant visual resemblance to the secret, which could lead to the unintentional revelation of structural and semantic information. InvCISD overcomes this challenge by integrating the latent representations of the secret and an unrelated reference image through an invertible network known as LIMNet. The framework involves training LIMNet in diffusion latent space, followed by comprehensive fine-tuning. This paper is a preprint and has been flagged as a cross-type submission.