Hidden Space Vector Coding-Based Generative Image Steganography With Diffusion Model

Recently, generative image steganography that hides messages in generated images has been a promising technique to achieve effective secret communication, due to its strong anti-steganalysis ability. However, existing generative image steganography still performs poorly in the image generation quality and hiding capacity. To resolve these shortcomings, this work proposes a Diffusion-based Generative Image Steganography (DGIS) approach, which includes a Diffusion Model (DM)-based stego-image generator, an INN-based secret autoencoder, and a UNet-based secret extractor. In this approach, the secret message is mapped to a hidden space vector to control the process of stego-image generation. Experimental results demonstrate that the proposed DGIS significantly improves image generation quality and message hiding capability.