Abstract
While encryption can ensure secure message exchange, the use of encryption can draw attention to the communication because encrypted content is not human readable. While text steganography can embed hidden messages in human-readable text, it is vulnerable to disclosure to unintended recipients. This disclosure describes steganography techniques that leverage machine learning techniques, specifically generative pre-trained transformer (GPT) type large language model (LLM) to generate human-readable natural language text that seamlessly conceals a different message within the text. The use of GPT can enhance the fluency and coherence of the natural language text, thus making it difficult to detect the secret message hidden within it. Per the techniques, logits (unnormalized predictions) generated by a language model are employed to encode a hidden message into unrelated human-readable text. Encoding is achieved by transforming tokenized natural language text into logits and applying ordering and optional additional cryptographic scrambling.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
n/a, "Text Steganography Using Large Language Model", Technical Disclosure Commons, (September 17, 2024)
https://www.tdcommons.org/dpubs_series/7366