Premise
Explanation 1:
The encoder is sending a secret message past a middle man to the decoder
The message is embedded within LLM text
We quantify how well we did on:
Sneakiness i.e. can the middle man notice changes to the secret message
Information Embedding Rate i.e. how much information can we place in each token
Explanation 2:
We have person a, b, c
Wlog, player a is sending a hidden message to player b through some cover medium
Player c inspects the cover medium for any distortions to detect the existence of the hidden message
In our context, the cover medium is LLM text.