Robust Generative Steganography via Intermediate State Normal Distribution

Generative steganography (GS) distinctively generates stego images directly from secret data, eliminating the need for traditional cover images. However, contemporary GS methods often encounter criticism due to their low extraction accuracy and compromised robustness. To address this limitation, we present ISND-Stego, a reliable and robust generative image steganography framework utilizing denoising diffusion implicit models. ISND-Stego capitalizes on the probability distribution of randomly sampled noise and intermediate state normal distribution in the diffusion model’s reverse process, embedding secrets through forward mapping. We meticulously designed and evaluated three bidirectional encryption schemes to guarantee both the extraction accuracy and the quality of stego images. The mapped stego noise follows the same probability distribution as standard sampling noise. Ultimately, the receiver retrieves the secret information from the intermediate state image via the reverse process. Comprehensive experiments demonstrate that ISND-Stego exceeds the capabilities of current state-of-the-art methods in terms of extraction accuracy, anti-detection, and image quality.