A Recipe for Steganogravy
5 days ago
- #data-privacy
- #AI-security
- #steganography
- AI scrapers and government agencies collect online data, making it risky to share information.
- Recipe-blog-encoding is a Python tool that uses neural linguistic steganography to hide data in recipe blog introductions.
- The tool encodes secret messages into natural language text using a shared prompt and model, like Qwen3-32B-Q4_K_M.gguf.
- Decoding requires the exact same model, quantization, and prompt; any mismatch produces garbage output.
- The implementation is based on arithmetic coding steganography, converting secrets into intervals for token selection.
- Challenges include low efficiency (<2-3 bits/token), tokenization mismatches, and early end-of-sequence termination.
- Security relies on the prompt as a shared key, but it's leaky, suitable for low-threat scenarios like passing notes.
- Only local LLMs are implemented, though remote APIs could work with determinism and logits.
- The tool is available on GitHub for installation and use, with a fix to ensure token round-trip consistency.