STEAD: Robust Provably Secure Linguistic Steganography Based on Diffusion Language Models

STEAD proposes a robust provably secure steganography method based on diffusion language models, overcoming the shortcomings of steganographic text from autoregressive models, which suffers from error propagation after tampering and cannot resist active tampering. It exploits the partial parallel generation characteristic of diffusion language models to select robust embedding positions and combines this with error-correcting codes; during extraction, it introduces pseudorandom error correction and neighborhood search error correction. The method can resist the token ambiguity of steganographic text segmentation and, to a certain extent, resist token-level insertion, deletion, and substitution attacks.