safety security
Privacy Audits Were Selecting Their Own False Alarms
Correcting membership-inference selection bias restored violations to the claimed confidence rate.
Summary
Correcting membership-inference selection bias restored violations to the claimed confidence rate.
The authors identify a selection bias in statistical audits of differential privacy that can overstate leakage lower bounds. Reanalysis found uncorrected audits crossing theoretical bounds up to four times more often than their confidence level implied. Standard corrections returned violations to no more than the nominal rate.
Why it matters
Correcting membership-inference selection bias restored violations to the claimed confidence rate.
Limits and context
No additional limitation was separately recorded.
Key claims
Correcting membership-inference selection bias restored violations to the claimed confidence rate.
Evidence: source-2026-10-21-019
Sources
- arXiv preprint 2610.27021, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.