Per-Record Privacy and Its Application to Heavy-Tailed Economic Data
summary
Claims a new formal guarantee called Per-Record Differential Privacy and demonstrates it by splitting establishments into pieces and applying standard DP with group-privacy accounting (example payrolls 1,000 2,000 5,000 and 4,000,000). Sherwin‑Williams is modeled as 4 pieces, JCPenney as 12, yielding per-record privacy loss = pieces × ε. A demonstration County Business Patterns product is deployed for ~8 million establishments and ~800 million tabulations; with noise SD ≈500 utility improves for cells with ≥10 establishments and 99% receive strong protection, while very large firms remain visible.
AI-written summary · about
chapters (8)
- 0:00Per-record privacy research
- 1:14Differential privacy with skewed data
- 2:42Per-establishment privacy loss
- 4:07Group privacy for large records
- 6:18Per-record privacy guarantees
- 7:01Privacy versus utility trade-offs
- 8:26Measuring protection in practice
- 10:17Future challenges
related talks
19mPractical Considerations for Differential Privacy
42mDifferential Privacy
27mPrivacy-preserving statistical analyses
20mShare data without revealing personal information
45mPractical Experience with Making Synthetic Data Differentially Private
20mDesigning for User Privacy: Integrating Differential Privacy into Ad Measurement...
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
27mYou've got trust issues, we've got solutions: Differential Privacy