Proving Model Equality in Zero-Knowledge
SPAR Spring 2026: Using ZKPs and challenge-response protocols to prove a deployed model matches a reference without revealing weights.
Problem. When an AI model is deployed as a service, users have no guarantee that the model they interact with is the same one that was audited or certified. Model providers could swap in a cheaper or modified model without detection.
Approach. We develop zero-knowledge proof systems and challenge-response protocols that allow a model operator to prove that the deployed model matches a reference model — without revealing the model weights to the verifier.
A SPAR Spring 2026 mentorship project.