How InferTrust Works

InferTrust creates cryptographic proof at the moment of AI inference. Learn the four-step process: capture, sign, seal, and verify. Sub-millisecond overhead, offline-capable, tamper-evident.

The Four-Step Process

1. Capture: InferTrust intercepts the inference at the execution boundary and captures model version hash, input feature hash, confidence score, decision action, active policy hash, and timestamp
2. Sign: A device-bound cryptographic key (TPM, Secure Enclave, or HSM) signs the record on the local device before any network transmission
3. Seal: The signed record is appended to a tamper-evident, append-only log where each entry chains to the previous entry hash
4. Verify: Any authorized party can independently verify any decision record using standard cryptographic libraries

What Gets Recorded

• Model Version ID (MID): SHA-256 of model binary
• Confidence Score (CS): float, used for confidence gating
• Policy Version Hash (PVH): SHA-256 of policy ruleset
• Input Feature Hash (IFH): SHA-256 of input tensor/features
• Timestamp (TS): Monotonic clock, cryptographically bound
• Decision Action (DA): APPROVE, ESCALATE, DENY, FLAG
• Sequence ID (SEQ): Strictly monotonic uint64