Manage training data in artificial intelligence

A system for verifying provenance of trained artificial intelligence models, comprising:

• a secure hardware processor;
• and
• a memory comprising instructions stored thereon, which when executed by the one or more processors, causes the one or more processors to perform:
• receiving, at a distributed ledger, training data from one or more data sources;
• issuing a request to train one or more models using the training data;
• receiving, from owners of the training data, approval of the request via a smart contract;
• executing a first transaction on the distributed ledger compensating owners of the training data for providing their data for the training of the one or more models, the first transaction serving as record on the distributed ledger of data inputs used to generate one or more trained models and granting access to the training data upon completion of a payment transaction to distributed ledger addresses of the owners;
• training the one or more models using the training data in response to the approval for access to the training data;
• generating a cryptographic certificate by executing back propagation for training the one or more models, the cryptographic certificate attesting that a back propagation routine was executed within a trusted execution environment of the secure hardware processor and binds (i) a set of training inputs corresponding to the training data used to train the one or more models, (ii) parameters of the one or more models resulting from the training, and (iii) an identifier and configuration of the secure hardware processor;
• generating, using the trained one or more models, a response to a query from a user based on at least a subset of the training data;
• executing a second transaction on the distributed ledger compensating owners of the subset of the training data for leveraging their data to generate the response to the query;
• verifying, using the cryptographic certificate, that the response was generated by a model trained within the trusted execution environment and without use of unapproved or unverified datasets;
• and
• tracking, via the distributed ledger, each use of the training data and each inference generated using the one or more models including model updates and/or query response generation.

A non-transitory computer-readable storage medium comprising instructions stored thereon, which when executed by one or more processors, cause the one or more processors to perform operations for verifying provenance of trained artificial intelligence models, comprising:

• receiving, at a distributed ledger, training data from one or more data sources;
• issuing, at a secure hardware processor executing a set of instruction codes, a request to train one or more models using the training data;
• receiving, from owners of the training data, approval of the request via a smart contract;
• executing a first transaction on the distributed ledger compensating owners of the training data for providing their data for the training of the one or more models, the first transaction serving as record on the distributed ledger of data inputs used to generate one or more trained models and granting access to the training data upon completion of a payment transaction to distributed ledger addresses of the owners;
• training, at the secure hardware processor, the one or more models using the training data based on the approval for access to the training data;
• generating, by the secure hardware processor, a cryptographic certificate by executing, at the secure hardware processor, back propagation for training the one or more models, the cryptographic certificate attesting that a back propagation routine was executed within a trusted execution environment of the secure hardware processor and binds (i) a set of training inputs corresponding to the training data used to train the one or more models, (ii) parameters of the one or more models resulting from the training, and (iii) an identifier and configuration of the secure hardware processor;
• generating, using the trained one or more models, a response to a query from a user based on at least a subset of the training data;
• executing a second transaction on the distributed ledger compensating owners of the subset of the training data for leveraging their data to generate the response to the query;
• verifying, using the cryptographic certificate, that the response was generated by a model trained within the trusted execution environment and without use of unapproved or unverified datasets;
• providing the response to the user based on a completion of the second transaction, wherein submitting the second transaction to the distributed ledger triggers residual payments from an account associated with the models to each owner corresponding to the subset of the training data;
• and
• tracking, via the distributed ledger, each use of the training data and each inference generated using the one or more models including model updates and/or query response generation.