SemanticZip: Improving Text Steganographic Capacity via LLM-Guided Semantic Lossless Compression

Existing generative text steganography methods primarily focus on optimizing the embedding stage while neglecting semantic redundancy mining and efficient coding in the preprocessing phase, which severely limits improvements in steganographic capacity. To address this issue, we propose a large language model-based framework for capacity enhancement, comprising an Adaptive Semantic Lossless Extraction module and a Probability-guided Entropy Adaptive Coding module. The former innovatively establishes a semantic lossless evaluation system to maximize the compression space of redundant text while ensuring strict semantic consistency. The latter integrates arithmetic coding with LLM probability distribution-based token-rank mapping and introduces entropy constraints, achieving highly efficient bitstream compression that approaches the Shannon entropy limit. Experimental results demonstrate that when integrated with mainstream generative steganographic algorithms, our framework increases steganographic capacity by 30.25\% and reduces processing time by 8.23 seconds on average. This method significantly improves transmission efficiency and minimizes transmission frequency, thereby further enhancing steganographic covertness and resistance to steganalysis.