Improving Safety with Privacy: An Exploration of Privacy...
summary
Argues that encryption, confidential computing and hardware-backed attestation let cloud inference run safety checks without routine human access by publishing approved binary measurements in a public append-only log and using per-client nonces. Gives concrete defenses: hash-matching stopped CSAM uploads, chained-request detection exposed Indonesian scam funnels, cross-account correlation revealed incremental malware builds, and cites the NYT demand for 20 million chats as the legal risk. Describes architecture for online/offline classifiers and an end-to-end no operator access guarantee, and flags open problems: debugging, distribution shift, metadata leakage, side channels, and attestation compromise.
AI-written summary · about
chapters (10)
- 0:00Privacy and safety in AI
- 1:02AI safety beyond alignment
- 4:12Detecting misuse in requests
- 9:54Defense in depth for AI safety
- 13:49Privacy versus safety in AI
- 16:36Hardware-backed privacy for AI
- 19:22Automating safety in private AI
- 21:28Debugging private AI services
- 23:33Private AI safety challenges
- 24:16Future directions and hiring
related talks
36mSecurity for AI and Multi-Party Collaboration with Confidential Computing (Oh, and...
01h 03mHow I Learned to Stop Worrying and Trust the Cloud
01h 00mSecure Computing and Hands-on Privacy By Design
48mSecurity today: change your habits and make it safer and efficient
29mPrivacy-preserving methods: Building secure projects
22mPrivacy for Safety - OPSEC When the Threat is in the Home
24mSafeguarding Sensitive Data Access At Scale with Privacy-Enhancing Technologies
26mUse case deep dive: Secure and privacy-preserving cyber threat intelligence exchange