COVER is a demo of a perfect-secrecy steganography algorithm that embeds data in LLM generated text without affecting the distribution of tokens.
In addition to preserving the LLM's natural distribution, this method is maximally efficient from the perspective of information theory.
This algorithm may be used to transmit data in secret or to mark LLM text with an invisible signature.
This implementation uses the tiny SmolLM2-135M language model; a larger model will improve the quality of text.
You can learn more about this algorithm by checking out my post here.
Context: The most well-known European staple food originally brought from the Americas is
Message: Look for a door marked with a purple lotus, knock twice and call