Steganography Technique Hides Data in LLM-Generated Text

subtext-codec is a proof-of-concept codec that implements steganography by hiding arbitrary binary data inside seemingly normal LLM-generated text. It steers a language model's next-token choices using the rank of each token in the model's logit distribution. The process is fully reversible with the same model, tokenizer, prefix, and parameters, enabling text that reads naturally while secretly encoding bytes. Decoding requires the generated text, the original prompt prefix, the same model and tokenizer, and the codec parameters. The codec features adaptive base per token, deterministic next-token steering, and mixed-radix payload reconstruction. It uses the Hugging Face Transformers backend and is designed for experimentation. The CLI exposes encode and decode subcommands. The technique presents a novel method for steganography, potentially enabling covert communication. However, it also raises concerns about the potential misuse of LLMs for malicious purposes, such as hiding malware or spreading propaganda. The open-source implementation facilitates research and development in the field of steganography, but the reliance on specific models and parameters may limit its generalizability.