Crypto + ML 2
Talkat Cryptography 10 Years Later: Obfuscation, Proof Systems, and Secure Computation Boot Camp 2025 · 2025-05-19..05-23·
summary
Concrete demos and limits: practical attacks on watermarking, model extraction, and fine-tuning protections, plus what cryptography could formalize. Shows token‑level watermark ideas and the semantic watermark robustness gap, a ReLU hyperplane method that recovers hidden neurons by walking decision boundaries, and a logit‑matrix attack that exposes an internal dimension via the logit matrix singular spectrum (tested on publicly billed APIs for a few dollars before providers tightened outputs). Also reports simple breaks of “unfine‑tunable” defenses and raises hardware DRM and auditable‑training questions needing crypto definitions or impossibility proofs.
AI-written summary · about
chapters (10)
- 0:00Introduction to speaker
- 1:27Classical crypto approaches
- 5:01Watermarking LLM outputs
- 10:40Model stealing techniques
- 18:10Small model weight extraction
- 29:56Stealing large models
- 38:06Stealing defenses and watermarks
- 41:55Modern model protections
- 48:37Compute governance
- 52:40Auditable training and closing
related talks
51mMultilinear Maps and Attacks II
44mThe Cryptographers' Panel
22mPrivacy-preserving Computation in Plain Sight
36mSmall Leaks, Billions Of Dollars: Practical Cryptographic Exploits That Undermine Crypto Wallets
52mMachine Learning over Encrypted Data with Fully Homomorphic Encryption
42mBreak Crypto like a Pro!
33mThe Promise of Trustworthy AI
49mPanel Discussion: Computational Assumptions