Risk-Equalized Differentially Private Synthetic Data: Protecting Outliers by Controlling Record-Level Influence
Amir Asiaee, Chao Yan, Zachary B. Abrams, Bradley A. Malin · 2026
Read Risk-Equalized Differentially Private Synthetic Data: Protecting Outliers by Controlling Record-Level Influence by Amir Asiaee, Chao Yan, Zachary B. Abrams, Bradley A. Malin (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-10232.txt.
Description
Some records need more protection than others. Take a patient with a rare illness mix or a deal with unusual traits, standing out from the crowd. A 2026 paper led by Amir Asiaee, with three co-authors, observes that membership inference attacks succeed much more often against such outliers, mainly under modest privacy budgets with auxiliary information. The team introduces risk-equalized DP synthesis, which prioritizes high-risk records by shrinking their influence on the learned generator.
A small privacy budget first gauges each record's outlierness; then a DP learning step weights each record in inverse proportion to its risk score. Under Gaussian mechanisms, privacy loss tracks influence on the output, so shrinking outliers' contributions tightens per-instance bounds exactly where protection matters most. Experiments with controlled outlier injection show risk-weighting substantially cuts membership inference success against high-outlierness records, and ablations confirm targeting, not random downweighting, drives the gain.
Catalog record
| CREATOR | Amir Asiaee, Chao Yan, Zachary B. Abrams, Bradley A. Malin |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 51 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-10232