Abstract
Systems and methods are provided for watermarking via unique synthetic datasets. A candidate model may be trained with a poison dataset including synthetic information. The synthetic information may relate to a fictitious place, person, or scenario, for example. A set of one or more inference questions associated with the poison dataset may be determined. The candidate model may be tested based on the set of one or more inference questions. A confidence may also be generated. The confidence may be associated with one or more identification metrics associated with the candidate model.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
"FICTITIOUS DATA-BASED WATERMARKING SYSTEM", Technical Disclosure Commons, (July 28, 2025)
https://www.tdcommons.org/dpubs_series/8400