Zero-Knowledge Machine Learning (ZKML) applies zero-knowledge proofs to machine learning so that one party can certify a training, testing, or inference result was actually produced by the claimed computation — without revealing the sensitive data or the proprietary model parameters behind it. That combination of privacy and verifiability is what separates ZKML from the other privacy-preserving techniques used in AI, most of which protect data but cannot prove that a computation was performed correctly. A survey consolidating ZKML research published from June 2017 to August 2025 organizes the field around three core tasks: verifiable training, verifiable testing, and verifiable inference. The cost of that guarantee is equally clear in the same literature: high proving cost, limited circuit expressiveness, and deployment complexity remain the field's main bottlenecks.
Zero-Knowledge Machine Learning: Verifiable AI That Hides Both Data and Model Weights
Zero-Knowledge Machine Learning (ZKML) leverages zero-knowledge proofs to certify AI computation results, offering both privacy for sensitive data and model parameters, and verifiability of the computation itself. This unique combination distinguishes ZKML from other privacy-preserving AI techniques, enabling auditable and trustworthy machine learning applications.
November 19, 2026 · 5 min read





