Ball Differential Privacy: Mitigating Data Reconstruction with Less Noise

Cet évènement est passé.

Quand

10 novembre 2026    
16 h 00 min - 18 h 00 min

Où

Évènement en ligne

Premedical, Inria – Nirupam Gupta (University of Copenhagen)

Saint-Priest Campus, Building 5, room 01/124

Differential privacy (DP) tries to hide participation, i.e., whether a record was used in training or not. In many settings, however, participation is not the sensitive fact; the content of the record is. This makes reconstruction a more practical threat: an informed adversary recovering a private record from a released model. DP does defend against reconstruction, but at a steep price. Its noise is calibrated to hide every conceivable swap of a record, even with records nowhere near anything the attacker finds plausible. This talk presents Ball-DP, which removes this excess noise. Ball-DP keeps DP’s indistinguishability argument intact but aims it where reconstruction happens: among records within a radius r of each other. The radius becomes an explicit privacy parameter, noise scales with r rather than the entire data domain, and standard DP is recovered when r is large enough to cover all plausible records. We obtain bounds on an attacker’s reconstruction success under Ball-DP and audit them with MAP reconstruction attack. The outcome is a principled defense against reconstruction threat, with significant accuracy gains over standard DP, most pronounced in the high-privacy regime.

[ Visio ]