Model Inference Attestation: A Technical Guide
Technical Guide

Model Inference Attestation

Model inference attestation is the runtime verification that a specific inference request was executed by an expected, unmodified model instance under an approved configuration. It relies on cryptographic evidence generated inside a trusted execution environment and validated against known reference values, distinguishing it from provenance tracking, which records model lineage rather than runtime execution integrity.

What Model Inference Attestation Verifies

Model inference attestation addresses a specific gap in production AI systems: the inability to confirm, at the moment an inference request is executed, that the responding system is running the expected model version under the expected configuration and has not been substituted or tampered with. This is distinct from confirming that a model was trained correctly or that its lineage is documented. Attestation is a runtime claim, not a design-time or artifact-level claim.

In practice, attestation must bind together three separate integrity claims: platform trust (the underlying hardware or virtual machine has not been compromised), model artifact identity (the exact weights, configuration, and tokenizer loaded match an approved reference), and runtime configuration (decoding parameters, safety filters, and other settings match policy). A verification that covers only one of these claims, for example hardware integrity without model identity binding, does not constitute complete inference attestation.

How Attestation Evidence Is Generated and Verified

Attestation builds on established remote attestation architecture rather than a novel AI-specific protocol. IETF RFC 9334 defines the RATS architecture, which separates three roles: the Attester, the system being verified, which produces evidence; the Verifier, which validates that evidence against policy and known-good reference values; and the Relying Party, which consumes the verification result to make an access or trust decision.

In hardware, this evidence typically originates from a Trusted Platform Module or a trusted execution environment such as Intel TDX or AMD SEV-SNP, which generate signed measurements of code and configuration that cannot be forged