[arXiv]score: 0.52
Tight Closed-Form DP-SGD Privacy Bounds for Random Shuffling Subsampling
May 26, 2026
Using the Berry-Esseen theorem within the f-DP framework, this work derives tight, closed-form trade-off function bounds for DP-SGD with random shuffling subsampling, achieving near-ideal privacy (δ=1/100, σ=1) with ~1000 samples per round.
cs.LGcs.CR
HOW THIS AFFECTS YOU
●
builderYou can use these transparent bounds to set concrete, auditable privacy parameters for DP training pipelines without relying on opaque numerical computations.
●
researcherClosed-form bounds for random shuffling DP-SGD are more interpretable than Poisson subsampling's implicit formulas, enabling cleaner theoretical analysis.
●
policyTighter, interpretable privacy accounting for DP-SGD strengthens the auditability of differentially private ML systems under regulatory scrutiny.