Opening Keynote
summary
Many common datasets lack provenance and long-term consent; a concrete fix is explicit data-as-contract records (version tables that link processors, consent status and expirations) plus runtime provenance queries. Practical tools discussed include the open-source Claro consent tool (English/German), on-device differential privacy like Apple’s emoji pipeline with local noise and a privacy budget, membership-inference attacks that reach 70–95% accuracy, and privacy-preserving training via PATE with ~5,000 teacher models (≈10–15% accuracy loss). Secure enclaves (e.g., Intel SGX) enable private training.
AI-written summary · about
chapters (13)
- 0:00Opening keynote introduction
- 3:14Origin of machine learning datasets
- 7:11Ethical concerns in data collection
- 11:56Data contracts and consent
- 17:16Meaningful consent implementation
- 19:45Tracking data provenance and consent
- 23:20Data privacy and pseudonymization
- 26:48Differential privacy in machine learning
- 31:50Privacy guarantees in ML models
- 38:52Differentially private training methods
- 43:09Secure computing for ML training
- 45:17Summary and call to action
- 48:30Q&A session
related talks
11mA New Model for International, Privacy-Preserving Data Science
39mKeynote 3: How did you know? Explaining Black Box Model Predictions in Python
22mLet’s make the world a better place: Python for Social Good
55mLightning Talks - Day 1
09mExpanding Differentially Private Solutions: A Python Case Study
06mBackstage interview
07mBackstage interview
29mOpen Source, Python Based Tools For Data Privacy