Dynamic privacy loss management: The innovation of α-confidence differential privacy filters and odometers

Abstract

In differential privacy, query perturbation plays a crucial role in ensuring privacy. Managing multiple queries requires amalgamating the privacy loss from each round to determine the overall loss. However, existing composition theorems are limited to specific mechanisms and do not account for query independence. This paper introduces α-Confidence Differential Privacy, a novel mechanism that calculates the minimum privacy loss based on database query events, overcoming mechanism-specific constraints. We formulate a composition theorem for α-Confidence Differential Privacy and develop dynamic privacy loss control through filter and odometer algorithms. Additionally, to address concurrent queries across different mechanisms, we establish an interactive composition theorem and implement corresponding interactive filters and odometers. The simulated experiments demonstrate the efficacy and practicality of our approach in the realm of differential privacy.