This is an official implementation of the paper "Addressing Segmentation Ambiguity in Neural Linguistic Steganography" (Accepted at AACL-IJCNLP2022).
$ git clone https://github.com/jumon/himitsu.git
$ pip install .
The following command encodes a secret message (a string of 0s and 1s) into cover text.
Given a language parameter (en, ja, ru, default: en), the corresponding GPT-2 model is used for encoding.
The parameter prompt is the context used to generate the cover text.
For other parameters, see the encode.py script.
$ python encode.py "010101011111010101101010" --prompt "Hi Bob." --language "en"
It would appear there's nothing new.
To decode the secret message from the cover text, use the decode.py script.
Remember to add leading spaces to the cover text if there are any.
The optional parameters have the same meaning as in the encode.py script and the same parameter values need to be used.
$ python decode.py " It would appear there's nothing new." --prompt "Hi Bob." --language "en"
01010101111101010110101000
The decoded secret message should be the same as the original secret message except for a few trailing 0s.