About
Hide secret messages inside naturally-generated text using next-token probability steering. Encodes information into LLM outputs by biasing token selection at inference time.
Key Highlights
- ▸ Implemented novel steganographic encoding scheme using LLM next-token probability distributions
- ▸ Encodes arbitrary bit strings into fluent natural language with no perceptible signal
Technologies
llm nlp steganography research python
Roles
llm-engineer researcher