Understand how InferTrust™ uses input feature hashing to prove what data an AI model evaluated without storing the raw sensitive data.
The Input Feature Hash (IFH) is a field in every InferTrust™ (Patent Pending) decision record that provides a cryptographic fingerprint of the data the AI model evaluated when making its decision. It proves what the model saw without storing the raw data itself.
In regulated industries, organizations need to prove what data an AI model used to make a decision. But storing copies of that data in decision records creates privacy and compliance problems, especially when the input includes protected health information (PHI), personally identifiable information (PII), or financial data. Simply referencing the source data is insufficient because the source could be modified after the decision.
At inference time, InferTrust™ computes a SHA-256 hash of the input features presented to the model. This hash is included in the signed decision record. If anyone later questions what data the model evaluated, the organization can retrieve the original data and recompute the hash. If the hashes match, it proves the data is unchanged. If they do not match, it proves the data has been modified since the decision was made.
Because the IFH is a one-way cryptographic hash, it cannot be reversed to reconstruct the original data. This means InferTrust™'s audit trail proves data provenance without creating additional copies of sensitive information, satisfying both auditability and data minimization requirements simultaneously.