Ball Differential Privacy: How to Mitigate Data Reconstruction with Less Noise
The paper introduces Ball-DP, a privacy mechanism that enforces differential privacy only over single-record substitutions within a specified radius 'r' of an embedding point. This local calibration allows for less noise than standard differential privacy while maintaining high model accuracy. It provides reconstruction-robustness certificates, Ball-ReRo, which quantify the attacker's success against this localized protection.