Calibrating Kernel Privacy Auditors: Known-Private Controls and the Cost of Validity

An auditor should rarely report a privacy violation when the mechanism satisfies the claimed guarantee. We study this requirement for the privacy application of regularized f -divergence kernel tests. The source paper's main kernel-test result, the DrMMD curve in Figure 3, reproduces in our experiment; our concern is the interpretation of its hockey-stick permutation test as a privacy auditor. Permutation calibration controls error under equal output distributions. Differential privacy allows unequal distributions, even at its boundary. On 14 known-private Laplace cells, the released auditor rejects 1,675/2,800 times and an estimator-corrected version rejects 2,137/2,800 times. The corrected version exceeds the nominal level in every cell. The effect also occurs in other mechanism families. Histogram examples establish limiting false-accusation probabilities of one half or one, depending on partition complexity; a corresponding Gaussian-kernel theorem remains open. A valid split auditor makes 1/2,800 false accusations but has limited power near the boundary. Combining held-out betting evidence across two folds improves power from 7% to 19% at ρ = 1.25 and from 62% to 93% at ρ = 1.5, with ϵ = 1 and n = 1000 per law. These power estimates describe the tested grid. The Gaussian-kernel results are empirical.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24369555
Primary Topic
Privacy-Preserving Technologies in Data
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Calibrating Kernel Privacy Auditors: Known-Private Controls and the Cost of Validity

S M Abdulla Al Mamun
Zenodo (CERN European Organization for Nuclear Research)
Privacy-Preserving Technologies in Data
preprint

Calibrating Kernel Privacy Auditors: Known-Private Controls and the Cost of Validity

S M Abdulla Al Mamun
preprint en

Abstract

An auditor should rarely report a privacy violation when the mechanism satisfies the claimed guarantee. We study this requirement for the privacy application of regularized f -divergence kernel tests. The source paper's main kernel-test result, the DrMMD curve in Figure 3, reproduces in our experiment; our concern is the interpretation of its hockey-stick permutation test as a privacy auditor. Permutation calibration controls error under equal output distributions. Differential privacy allows unequal distributions, even at its boundary. On 14 known-private Laplace cells, the released auditor rejects 1,675/2,800 times and an estimator-corrected version rejects 2,137/2,800 times. The corrected version exceeds the nominal level in every cell. The effect also occurs in other mechanism families. Histogram examples establish limiting false-accusation probabilities of one half or one, depending on partition complexity; a corresponding Gaussian-kernel theorem remains open. A valid split auditor makes 1/2,800 false accusations but has limited power near the boundary. Combining held-out betting evidence across two folds improves power from 7% to 19% at ρ = 1.25 and from 62% to 93% at ρ = 1.5, with ϵ = 1 and n = 1000 per law. These power estimates describe the tested grid. The Gaussian-kernel results are empirical.

Zenodo (CERN European Organization for Nuclear Research)
University of Manchester (GB)
Privacy-Preserving Technologies in Data
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon