A Novel Approach to Differential Privacy with Alpha Divergence

2025-06-20

AI Summary

Conventional $(varepsilon,delta)$-differential privacy (DP) suffers from inherent limitations in adaptivity and fine-grained privacy control. Method: This paper proposes Alpha Differential Privacy (ADP), the first tunable DP framework systematically built upon $alpha$-divergence. ADP incorporates $alpha$-divergence into the privacy definition, enabling continuous, analytically tractable adjustment of privacy strength and dynamic privacy budget allocation. Contribution/Results: Theoretical analysis shows that ADP yields tighter privacy bounds than standard DP under small-to-moderate iteration counts. Empirical evaluation demonstrates ADP’s superiority—particularly under high-privacy requirements and limited iteration budgets—achieving 20%–45% stronger privacy protection across multiple benchmark tasks compared to baseline methods. Crucially, ADP establishes the first rigorous, $alpha$-divergence-based DP paradigm, unifying theoretical soundness with practical flexibility.

Technology Category

• Machine Learning: Privacy
• Search and Optimization: Mixed Discrete/Continuous Search
• Game Theory and Economic Paradigms: Adversarial Learning

Application Category

• Security and Privacy: Large-scale security measurements
• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Responsible Web: Data and user privacy-enhancing technologies for the Web

Abstract

As data-driven technologies advance swiftly, maintaining strong privacy measures becomes progressively difficult. Conventional $(epsilon, delta)$-differential privacy, while prevalent, exhibits limited adaptability for many applications. To mitigate these constraints, we present alpha differential privacy (ADP), an innovative privacy framework grounded in alpha divergence, which provides a more flexible assessment of privacy consumption. This study delineates the theoretical underpinnings of ADP and contrasts its performance with competing privacy frameworks across many scenarios. Empirical assessments demonstrate that ADP offers enhanced privacy guarantees in small to moderate iteration contexts, particularly where severe privacy requirements are necessary. The suggested method markedly improves privacy-preserving methods, providing a flexible solution for contemporary data analysis issues in a data-centric environment.

Problem

• Enhancing privacy in data-driven technologies with limited adaptability
• Introducing alpha differential privacy for flexible privacy assessment
• Improving privacy guarantees in small to moderate iteration contexts

Innovation

• ADP based on alpha divergence
• Flexible privacy consumption assessment
• Enhanced guarantees in strict contexts