Crypto + ML 1
Talkat Cryptography 10 Years Later: Obfuscation, Proof Systems, and Secure Computation Boot Camp 2025 · 2025-05-19..05-23·
summary
Shows that verification and attacks admit concrete crypto-style answers: PAC-verifiable for Fourier-sparse Boolean functions using Goldreich‑Levin-style checks with poly(t,1/ε,1/δ) sample/time bounds (improved in 2024); trains autoregressive models to emit accepting interactive‑proof transcripts (GKR-style proofs, Bezout/GCD prototypes) so outputs come with verifiable proofs; demonstrates undetectable black‑box backdoors (and white‑box CLWE/continuous‑LWE constructions in random‑Fourier‑feature models) and derives a time‑lock‑based jailbreak impossibility for fast input/output filtering.
AI-written summary · about
chapters (11)
- 0:00Introduction and motivation
- 3:50Defining trust in machine learning
- 8:40Privacy and verification challenges
- 17:26Verifying model accuracy
- 23:28Mathematical proofs of accuracy
- 28:22Self-proving generative models
- 35:46Transcript learning algorithms
- 41:40Future research directions
- 43:31Backdoors in machine learning
- 51:06Mitigation and white-box attacks
- 59:43Alignment and concluding remarks
related talks
57mThe Cryptographic Lens: Visions of our Past and Future
32mBlack box side-channel leakage verification using statistical approach
40mPost-Quantum
10mCryptography 10 Years Later: Obfuscation, Proof Systems, and Secure Computation
01h 08mTutorial on Proof Systems
57mEssentially Optimal Interactive Certificates in Linear Algebra
22mPrivacy-preserving Computation in Plain Sight
38mAdvisor-Verifier-Prover Games and the Hardness of Information Theoretic Cryptography