Abstract
This disclosure describes techniques to produce machine-generated content with strong guarantees of detectability. Sequential content is generated by a generative artificial intelligence tool. The inference procedure used to sample content from the model is extended to provide cryptographic guarantees that the content can be detected to be synthetic. The sampling step that occurs during the generation of sequential output from a generative model is augmented with provably secure steganographic techniques. At each generation step where candidate tokens are evaluated for inclusion in the output of the AI tool, tokens that do not satisfy a cryptographically defined condition relative to previous tokens are rejected. The cryptographic condition is designed to satisfy the formal threshold of detectability desired. Outputs generated as a result of this process are detectable as synthetic by computation of statistics on top of the count of sequence positions where the condition holds.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Alston, Aubrey, "Provably Detectable Sequence Generation from AI Models via Steganographic Embedding", Technical Disclosure Commons, (January 11, 2024)
https://www.tdcommons.org/dpubs_series/6595