Natural f-DP filters are invalid for adaptive composition, but a CLT-based GDP approximation gives tighter bounds than RDP for subsampled Gaussians when sampling rate is near 0 or 1.
Since the ∆-approximate µ-GDP guarantee is symmetric, the privacy guarantees above apply to both trade-off functions T[M 1:t(S),M 1:t(S−)]andT[M 1:t(S−),M 1:t(S)]
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CR 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
$f$-Differential Privacy Filters: Validity and Approximate Solutions
Natural f-DP filters are invalid for adaptive composition, but a CLT-based GDP approximation gives tighter bounds than RDP for subsampled Gaussians when sampling rate is near 0 or 1.