What every data scientist should know about data anonymization
summary
Numerical claims and concrete failures decide whether to watch: Netflix prize example shows ratings reidentification from two ratings plus dates; mobile-phone traces need only four spatio-temporal points → 95% to reidentify people. Methods tested: k-anonymity with generalization/suppression and experiments on the Adult dataset using rxd identification software show f1 stays similar up to K≈50 then collapses (half performance by K≈300) while many rows get suppressed. Alternative is query-based noise injection à la differential privacy.
AI-written summary · about
chapters (15)
- 0:00Data anonymization challenges
- 3:36Real-world anonymization failures
- 7:11Adversaries and privacy risks
- 10:26Quasi-identifiers and linkage attacks
- 14:30K-anonymity limitations
- 18:24Anonymization impact on utility
- 21:47High-dimensional data resistance
- 24:18Interactive query systems
- 26:28Practical anonymization steps
- 29:17Summary and recommendations
- 30:03Q&A: Noise addition and translation
- 31:53Q&A: Age demographics privacy
- 32:37Q&A: Database access and differential privacy
- 33:51Q&A: Rare cases suppression
- 34:36Q&A: Graph and map anonymization
related talks
15mHow We Can Save Anonymization
08mEvil Uses of Data
52mMaking Better Decisions with Data
32mYou've got trust issues, we've got solutions: Differential Privacy
20mWhy data privacy is critical for robust data management?
25mThe Importance of Synthetic Data in Data-Centric AI
11mMy $5MM Differential Privacy Visualizations
34mDimensionally reducing data squeezing out the good stuff