Verifiable inference is the capability to cryptographically prove that a specific machine-learning model produced a given output for a given input without requiring trust in the compute provider. Techniques include zero-knowledge proofs of model execution (zkML), trusted execution environments and optimistic verification, which let third parties audit results. It is essential for decentralised and trust-minimised AI compute markets where inference is outsourced.
Content
- Approaches trade off cost and assurance: zkML generates succinct proofs at heavy proving overhead, TEEs offer hardware attestation with weaker cryptographic guarantees, and optimistic schemes assume honesty with fraud-proof fallback. The goal is trust-minimised verification of outsourced AI computation.