Linguistic Steganography (LS) tasks aim to generate steganographic text (stego) based on secret information. Only authorized recipients can perceive the existence of secrets in the texts and extract them, thereby preserving privacy. However, the controllability of the stego generated by existing schemes is poor, and the stego is difficult to contain specific discourse characteristics such as style. As a result, the stego is easily detectable, compromising covert communication. To address these problems, this paper proposes LLsM, the first LS with the Large Language Model (LLM). We fine-tuned the LLaMA2 with a large-scale constructed dataset encompassing rich discourse characteristics, which enables the fine-tuned LLM to generate texts with specific discourse in a controllable manner. Then the discourse is used as guiding information and inputted into the fine-tuned LLM in the form of the Prompt together with secret. On this basis, the constructed candidate pool will be range encoded and use secret to determine the interval. The same prefix of this interval's beginning and ending is the secret embedded at this moment. Experiments show that LLsM performs superior to prevalent LS-task and related-task baselines regarding text quality, statistical analysis, discourse matching, and anti-steganalysis. In particular, LLsM's MAUVE matric surpasses some baselines by 70%-80%, and its anti-steganalysis performance is 30%-40% higher. Notably, we also present examples of longer stegos generated by LLsM, showing its potential superiority in long LS tasks.

翻译：语言学隐写术（LS）任务旨在根据秘密信息生成隐写文本。仅授权接收者能够感知文本中秘密的存在并提取之，从而保护隐私。然而，现有方案生成的隐写文本可控性较差，且难以具备特定的话语特征（如风格）。因此，隐写文本易被检测，危及隐蔽通信。为解决这些问题，本文提出LLsM——首个基于大语言模型（LLM）的隐写方案。我们使用涵盖丰富话语特征的大规模构建数据集对LLaMA2进行微调，使微调后的LLM能够以可控方式生成具有特定话语的文本。随后，将话语作为引导信息，与秘密一同以提示词（Prompt）形式输入微调后的LLM。在此基础上，构建的候选池将进行范围编码，并利用秘密确定区间。该区间起始与终止部分的共同前缀即为当前时刻嵌入的秘密。实验表明，LLsM在文本质量、统计分析、话语匹配及抗隐写分析方面均优于主流LS任务及相关任务基线。特别地，LLsM的MAUVE指标超越部分基线70%-80%，其抗隐写分析性能高出30%-40%。此外，我们还展示了LLsM生成的长文本隐写示例，证明了其在长文本LS任务中的潜在优势。