VDDP: Verifiable Distributed Differential Privacy under the Client-Server-Verifier Setup

2025-04-30

AI Summary

In distributed client-server-verifier architectures, malicious servers may tamper with noise distributions—e.g., via sampling bias or artificial correlations—thereby breaking differential privacy (DP) guarantees. To address this, we propose Verifiable Distributed Differential Privacy (VDDP), the first formal framework for verifying DP compliance in such settings. We establish a rigorous definition of VDDP and reveal a sufficient (but not necessary) connection between zero-knowledge proofs (ZKPs) and DP verifiability. We design two efficient mechanisms: (i) the Verifiable Discrete Laplace Mechanism (VDDLM), which accelerates proof generation by 4×10⁵× while incurring only 0.1–0.2× the error of baseline mechanisms; and (ii) Verifiable Randomized Response (VRR), which reduces both communication and verification overhead by 5000×. Collectively, VDDP significantly enhances the practicality and trustworthiness of deploying DP in distributed systems.

Technology Category

• Machine Learning: Distributed Machine Learning & Federated Learning
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization
• Search and Optimization: Distributed Search

Application Category

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

Abstract

Despite differential privacy (DP) often being considered the de facto standard for data privacy, its realization is vulnerable to unfaithful execution of its mechanisms by servers, especially in distributed settings. Specifically, servers may sample noise from incorrect distributions or generate correlated noise while appearing to follow established protocols. This work analyzes these malicious behaviors in a general differential privacy framework within a distributed client-server-verifier setup. To address these adversarial problems, we propose a novel definition called Verifiable Distributed Differential Privacy (VDDP) by incorporating additional verification mechanisms. We also explore the relationship between zero-knowledge proofs (ZKP) and DP, demonstrating that while ZKPs are sufficient for achieving DP under verifiability requirements, they are not necessary. Furthermore, we develop two novel and efficient mechanisms that satisfy VDDP: (1) the Verifiable Distributed Discrete Laplacian Mechanism (VDDLM), which offers up to a $4 imes 10^5$x improvement in proof generation efficiency with only 0.1-0.2x error compared to the previous state-of-the-art verifiable differentially private mechanism; (2) an improved solution to Verifiable Randomized Response (VRR) under local DP, a special case of VDDP, achieving up a reduction of up to 5000x in communication costs and the verifier's overhead.

Problem

• Addresses server vulnerabilities in distributed differential privacy execution
• Proposes Verifiable Distributed Differential Privacy (VDDP) with verification mechanisms
• Develops efficient VDDP mechanisms with improved performance metrics

Innovation

• Introduces Verifiable Distributed Differential Privacy (VDDP)
• Develops Verifiable Distributed Discrete Laplacian Mechanism (VDDLM)
• Improves Verifiable Randomized Response (VRR) efficiency