Tutorial: Privacy, Part I
summary
Covers rigorous limits and concrete attacks: models repeatedly emit training data (GPT-2 and diffusion memorization, census reconstructions), a classic result shows reconstruction from n linear queries when answers are accurate enough (alpha^2 n small), and high‑dimensional settings let attackers infer membership once features outnumber records so that membership inference when D>n with accuracy ~sqrt(D/n) succeeds even after per‑coordinate distortion; proposes differential‑privacy regimes and auditing as practical defenses.
AI-written summary · about
chapters (6)
- 0:00Privacy in machine learning
- 12:57Reconstruction attacks
- 27:03Membership inference attacks
- 50:14Privacy attack thresholds
- 1:06:43Open privacy research
- 1:16:40Summary
related talks
28mPrivacy preserving machine learning
21m"SECRETS ARE LIES, SHARING IS CARING, PRIVACY IS THEFT." - A Dive into Privacy Preserving
01h 43mPerspective
29mPrivacy-preserving methods: Building secure projects
31mIs your model private?
01h 12mModelling Privacy Threats of Medical Data and Health Data Spaces
43mGuarding the ML Galaxy: Beyond Accuracy to Privacy and Security
01h 14mStatistical Inference and Privacy, Part III