Zero knowledge machine learning with ezkl
summary
EZKL compiles PyTorch/ONNX models into zk circuits, quantizes weights to fixed‑point and generates a prover + EVM verifier so model inference can be proven and submitted on chain. Demo shows compiling, SRS/setup key generation, producing a proof and on‑chain verification; browser/WASM and Ledger submit tools exist. It supports private inputs/parameters, and claims quantizes to fixed‑point (denominator 256) and 5–6 convolutional layers <10s proving for typical models.
AI-written summary · about
chapters (18)
- 0:00Zero knowledge ML intro
- 1:04Physics engine for worlds
- 2:49ZK proofs as signatures
- 3:51Trustless AI on-chain
- 7:44How ZK proofs work
- 9:29ezkl for AI big picture
- 10:55Answering audience questions
- 12:31ezkl project demo setup
- 18:30Building custom circuits
- 21:02What the demo proves
- 22:48Private vs public data
- 24:53Private output data
- 26:44Deploying verifier contracts
- 27:50Calling verifier contracts
- 29:03Complex use cases
- 30:10AI in autonomous worlds
- 31:57JavaScript compatibility
- 33:08Closing
related talks
30mPractical Zero Knowledge Proofs
29mAccountable AI with ZKPs: Certifying Fairness and Explanations under model Confidentiality
25mHigh-Speed zkSNARKs without Trusted Setup
01h 26mPanel Discussion: The Challenges and Best Practices of Deploying Zero-Knowledge Proofs
22mDIZK: A Distributed Zero Knowledge Proof System
32mSoK: Private Programmability on Blockchain Systems
51mWhy you shouldn't trust me
20mThe Evolution of ZKProof.org