Does using data mean giving up privacy?
summary
Machine learning needs lots of data but privacy-preserving methods aim to let institutions share useful models without exposing individuals. homomorphic encryption exists but remains too slow for most use; multi-party computation is practical today for collaborative analytics (Boston/City of Boston salary study uses MPC with distributed shares and works albeit with high communication overhead). differential privacy introduces a tunable epsilon noise-utility tradeoff and the US Census adopts it; OpenMined provides a PyTorch Python library implementing MPC and differential privacy.
AI-written summary · about
chapters (22)
- 0:00Introduction
- 0:57Zombies eat brains
- 1:30Machine learning
- 1:56Private data
- 2:17Anonymization
- 2:54Serving a naive
- 4:01Traditional techniques
- 5:26Law briefs
- 5:37Identifying data
- 6:46Rei densification
- 7:56Strava example
- 9:43linkage attacks
- 12:40Research
- 13:01Differential Privacy
- 16:13The Census
- 16:58Controversy
- 18:56Lastbased Encryption
- 19:18Multiparty Computation
- 20:03Considerations
- 20:46Boston University
- 22:55Ongoing Research
- 23:45Questions
related talks
01h 19mWhen Data Can’t Move but Insight Must
20mShare data without revealing personal information
38mDemystifying Privacy Preserving Computing
29mPrivacy-preserving methods: Building secure projects
16mPrivacy Preserving AI
33mPractical Data Privacy
22mPrivacy-preserving Computation in Plain Sight
42mPrivacy Challenges in Machine Learning