Safety, Ethics & Governance · advanced · concept 176 of 176
Differential Privacy & ML Privacy
A mathematical guarantee that a model or statistic barely changes whether or not any single person's data is included, enforced by adding calibrated noise. The gold standard for training on sensitive data, used by the US Census, Apple, and Google, with a real accuracy cost to budget for.
Key terms
Privacy budget (epsilon)Noise injectionDP-SGDMembership inferenceAnonymization limits
Learn these first
Where you meet it in the real world
US Census releases, keyboard prediction telemetry, medical research on patient data
Videos
▶ Tutorial: Differential Privacy and Learning: The Tools, The Results, and The Frontier ↗
Microsoft Research · YouTube
▶ Differential Privacy for Growing Databases ↗
Microsoft Research · YouTube
Guides and articles
Courses, papers, and more