Stories about DP-SGD
1 related stories
Revisiting the Provable-Auditable Privacy Gap of DP-SGD
AI InsightThis paper revisits the gap between theoretical privacy upper bounds and empirical auditing lower bounds for DP-SGD. Previous auditing literature generally found DP-SGD's privacy bounds nearly tight; this work may reveal new analytical methods or a larger gap. The auditing consensus is no longer settled, and theoretical tightness claims need recalibration.Key TakeawayFrom auditing literature treating DP-SGD bounds as near-tight, to revisiting the provable-auditable gap.Why It MattersDP-SGD is the de facto private training method; bound tightness directly affects privacy budget setting and compliance decisions.Who's Affected- AI ResearchersNeed to reassess the gap between DP-SGD's theoretical bounds and auditing results.
- Privacy PractitionersA larger gap implies actual privacy guarantees may be stronger than theory.
- RegulatorsShifts in auditing conclusions affect credibility of compliance assessment methods.
What's NextWatch for new auditing lower bounds or tighter theoretical analyses, and whether findings generalize to other private training algorithms.Importance 72/100