Towards Provably Secure and Highly Robust Generative Image Steganography Leveraging Latent Diffusion Model

Generative image steganography has attracted significant at- tention due to its unparalleled resilience against steganaly- sis. However, current generative steganography methods still confront difficulties in terms of the lack of provable security guarantees under statistical analysis and vulnerability to real- world unknown channel attacks. To overcome these obsta- cles, this paper proposes a novel generative image steganog- raphy framework that leverages the Latent Diffusion Model (LDM). Notably, we have uncover a consistent trend: regard- less of whether an image has undergone attacks such as com- pression or noise addition, the sign of the values in its latent vector, encoded through LDM, remains unchanged. Capital- izing on this trend, we have devised a adaptive distribution- preserving mapping (ADPM) mechanism, capable of con- verting a secret message into a latent vector that follows stan- dard normal distribution in an adjustable way. Since both the secret latent vector and the latent vector randomly gen- erated during regular image generation follow the same dis- tribution, satisfying the optimal input conditions for the diffu- sion model, the proposed method can achieve provable secu- rity. The experimental results highlight the outstanding per- formance of our method in terms of robustness, security, ex- traction accuracy, and image quality.