Breaking the Gaussian Barrier: Residual-PAC Privacy for Automatic Privatization

Abstract: The Probably Approximately Correct (PAC) Privacy framework [1] provides a powerful instance-based methodology for certifying privacy in complex data-driven systems. However, existing PAC Privacy algorithms rely on a Gaussian mutual information upper bound. We show that this is in general too conservative: the upper bound obtained by these algorithms is tight if and only if the perturbed mechanism output is jointly Gaussian with independent Gaussian noise. To address the inefficiency inherent in the Gaussian-based approach, we introduce Residual PAC Privacy, an f-divergence-based measure that quantifies the privacy remaining after adversarial inference. When instantiated with Kullback-Leibler divergence, Residual-PAC Privacy is governed by conditional entropy. Moreover, we propose Stackelberg Residual-PAC (SR-PAC) privatization mechanisms for RPAC Privacy, a game-theoretic framework that selects optimal noise distributions through convex bilevel optimization. Our approach achieves tight privacy budget utilization for arbitrary data distributions. Moreover, it naturally composes under repeated mechanisms and provides provable privacy guarantees with higher statistical efficiency. Numerical experiments demonstrate that SR-PAC certifies the target privacy budget while consistently improving utility compared to existing methods.