Denoising Diffusion Probabilistic Steganography Based on Standardized Secret Noise

Xiang Zhang, Tianheng Song, Fei Peng, Ziwen He, Daoyong Fu, Bei Yuan, Zhangjie Fu

Abstract: Generative steganography based on diffusion model is a technique that directly uses secret information to generate a stego image by diffusion model. Due to its strong resistance to steganalysis, it has become a hotspot in current research. However, this technique faces the core issue of insufficient image quality, which stems from the mismatch between the stego noise distribution and standard Gaussian distribution, leading to cumulative noise prediction errors and deviation from the generation path. This paper designs a precise alignment strategy for noise distribution, matching the mean and covariance of the stego noise with the standard Gaussian distribution. This theoretically ensures the optimality of the reverse denoising path in the diffusion model and significantly reduces error accumulation. Furthermore, we leverage the text control capability of the pre-trained diffusion model, combined with semantic vectors and a cross attention mechanism, to dynamically adjust the generation of stego content, achieving high quality image synthesis. Experimental results show that, compared with existing methods, the proposed approach significantly improves both visual quality and robustness.

External IDs:doi:10.1109/lsp.2025.3624083