Post-quantum verifiable joint-watermarking for multiparty artificial intelligence models
AssigneeCircle Internet Group, Inc.
InventorVipin Singh Sehrawat
Life of this patent
Estimating the term…
Abstract
Methods, systems, and computer-readable storage media for receiving a prompt, providing a context window as a hash value at least partially based on a set of previously received tokens, determining a split parameter using a first key of a first party and a second key of a second party, the first key and the second key generated by a distributed key generation (DKG) process using a channel key, providing a green list of tokens and a red list of tokens using the split parameter, determining a watermark probability distribution that favors tokens in the green list of tokens over tokens in the red list of tokens, and returning watermarked digital content that is generated using the watermark probability distribution.
Claims10
A computer-implemented method for generating watermarked digital content using generative artificial intelligence (GAI), the method comprising:
- receiving a first prompt;
- providing a first context window as a hash value by processing a set of previously received tokens through a hash function;
- determining a first split parameter using a first key of a first party and a second key of a second party, the first key and the second key generated by a distributed key generation (DKG) process using a channel key, the first split parameter being used to assign tokens to one of a green list and a red list;
- providing a first green list of tokens and a first red list of tokens using the first split parameter;
- determining a first watermark probability distribution that is biased towards tokens in the first green list of tokens over tokens in the first red list of tokens, such that more tokens are selected from the green list of tokens than are selected from the red list of tokens in generating a watermark;
- and
- returning first watermarked digital content that is generated using the first watermark probability distribution.
Dependencies are inferred from the claim text.