Enhancing the communication reliability for generative image steganography with diffusion model

Abstract

Steganography, which conceals information within ordinary covers, can be applied in journalism, intelligence, and healthcare. It effectively avoids the risks of censorship, interception, and leakage, meeting the need for secure information transmission. Images generated by diffusion models are highly valued for their content authenticity and wide applicability, providing high-quality cover images for covert communication. However, existing methods face difficulties in generating stego images that conform to mainstream formats while ensuring the complete extraction of information. This may attract the attention of third parties and increase the risk of covert communication being detected and exposed. Therefore, we propose a diffusion model-based steganographic method to enhance communication reliability. The method first constructs a function that maps secret data to the latent vector space using non-continuous sub-intervals divided by the inverse cumulative distribution function and then generates an initial stego image using the diffusion model. Subsequently, an effective error detection mechanism is designed to address the potential loss of secret data during the quantization process based on the mapping rules and characteristics of the non-continuous sub-intervals. Finally, driven by the secret data, an adaptive quantization strategy is employed to iteratively correct the lost data based on the initial stego image and error detection information. Experimental results demonstrate that the proposed method can generate stego images in mainstream formats while demonstrating a consistently high extraction accuracy rate. Compared with methods that can ensure complete secret data extraction and use mainstream stego image formats, our method achieves state-of-the-art performance in embedding capacity.