Privacy-preserving Computation in Plain Sight
summary
Describes concrete patterns for model training that let an observer verify results without seeing data by combining zero-knowledge proofs, fully homomorphic encryption, commitments, garbled circuits and MPC. Shows prototypes: a sklearn→ZK conversion for an MLP, implementations in Noir and Circom, Rust/arkworks constraint optimizations, and verifiable-credential demos; two related grants pending. Claims privacy with cost: batched training or one-off proofs are practical today, real-time services require substantial optimization. Model parameter commitments plus execution proofs prevent IP or data leakage.
AI-written summary · about
chapters (11)
- 0:00Privacy-preserving computation
- 1:13Computation in plain sight
- 3:00Cloud data exposure risks
- 7:34Cryptographic tools for privacy
- 12:31Zero-knowledge proofs for cloud
- 14:39Homomorphic encryption training
- 16:23Secure data silo collaboration
- 17:28Private data intersection MPC
- 18:54Privacy versus performance
- 19:38ZK and HE research
- 21:28Q&A
related talks
42mPrivacy Challenges in Machine Learning
12mThe Limits of Technology in Achieving Privacy
29mPrivacy-preserving methods: Building secure projects
26mDoes using data mean giving up privacy?
28mZero Knowledge Proofs and Their Future Applications
19mPrivate AI: Building Trust Through Verifiable Computation
33mAn Introduction to Zero Knowledge Proofs
26mHow zero-knowledge proofs can be used to balance privacy and compliance