Perspective
summary
Machine attacks reconstruct records from published statistics—experiments report ~96–99% recovery and a simple birthday-style predicate singles out a random person with ~37% probability—so reconstruction attacks recovered ~99%. Linkage shows k-anonymity fails against linkage; noisy tables plus clever algorithms break many ad‑hoc defenses. Differential privacy uses calibrated noise (Gaussian, Laplace), an epsilon privacy budget with formal composition and resists those reconstructions; deployments include Apple, Google telemetry and Census work, while local models need massive participation and practical integration with crypto remains open.
AI-written summary · about
chapters (31)
- 0:00Intro
- 0:47Considerations in preparing this talk
- 2:17Layout
- 3:38Informational Privacy: The Problem
- 5:17When is informational privacy preserved?
- 13:27Reconstruction attacks [Dinur N '03]
- 20:34Reconstruction attacks Dinur N '03
- 26:14Attacks on traditional privacy enhancing techniques
- 26:38Reasons for optimism
- 30:40Reconstruction attacks are useful!
- 31:21What reconstruction attacks teach us?
- 32:25Progressing towards a definition
- 34:53A privacy desiderata
- 37:32A more realistic privacy desiderata
- 37:51Differential privacy [Dwork McSherry N Smith 86]
- 38:52New privacy concepts
- 40:06What makes a good privacy concept? (2)
- 44:03What makes a good privacy concept? (3)
- 45:58What makes a good privacy concept? (4)
- 50:19Differentially Private Computations
- 51:31What can be computed with differential Privacy?
- 52:17What differential privacy is not?
- 55:01U.S. Census Bureau
- 55:11GOOGLE
- 55:22Apple
- 55:39Some challenges
- 56:27A perspective talk
- 57:10Data Privacy: The Problem
- 57:51A computer scientist and a legal scholar meet at the Simons Institute
- 58:37What does "protecting individual privacy" mean?
- 59:18Legal approach
related talks
42mDifferential Privacy
01h 17mTutorial: Privacy, Part I
20mShare data without revealing personal information
33mPractical Data Privacy
27mThe ethical algorithm
29mPrivacy-preserving methods: Building secure projects
32mThe Dangerous Data Anonymization - Keynote
33mPrivacy of Decentralized Machine Learning