f f -Differential Privacy Filters: Validity and Approximate Solutions
Long Tran, Antti Koskela, Ossi Räisä, Antti Honkela · 2026
Read f f -Differential Privacy Filters: Validity and Approximate Solutions by Long Tran, Antti Koskela, Ossi Räisä, Antti Honkela (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-2602-06756.txt.
Description
Long Tran, Antti Koskela, Ossi Räisä, and Antti Honkela examine how to keep a privacy budget when a data analyst adapts on the fly. In differential privacy, a privacy filter is a stopping rule that halts the analysis before a total budget is overspent. The paper asks whether the natural filter built from f-differential privacy, a refined accounting framework, can be trusted in this fully adaptive setting.
The answer is no. Tracking the accumulating trade-off curve and stopping when it crosses the budget line turns out to be fundamentally invalid. That blocks the use of fast standard accounting tools. The authors pin down exactly when that natural filter could work. They also prove a central limit theorem for privacy loss under full adaptivity. Finally, they build a closed-form approximate filter for a common mechanism family that beats a popular alternative in the extremes.
Catalog record
| CREATOR | Long Tran, Antti Koskela, Ossi Räisä, Antti Honkela |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 149 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-06756