On Minimal Privacy Leakage Frequency in Differentially Private Decentralized Optimization
Year: 2026
Type: article
Abstract: Differential privacy has become a standard defense against gradient leakage attacks in distributed optimization, yet it consistently grapples with the security-utility trade-off. Enhancing security without compromising convergence performance remains a key challenge. In this paper, we address this issue by analyzing the algorithmic security through a novel metric called Privacy Leakage Frequency (PLF), which quantifies the degree of privacy exposure across optimization iterations. By noting that lower PLF implies stronger security, we firstly derive the lower bound on PLF for first-order distributed optimization algorithms as Ω(√κlog 1 ϵ), where κ is the condition number of local functions and ϵ is the convergence precision. To attain this bound, we propose DP-DOML, a differentially private decentralized stochastic algorithm based on a novel iterative mode termed K-gossip skip. Theoretically, we establish the convergence rate for DP-DOML under both general and strong convex objectives, confirming the achievement of minimal PLF. We further perform a local differential privacy analysis to confirm its enhanced security guarantees. Finally, extensive experiments on real-world datasets validate that DP-DOML outperforms existing methods in terms of convergence accuracy and privacy preservation.
Authors: Luqing Wang, Shaofu Yang, Wenying Xu, Wenwu Yu
Institutions: Southeast University, Purple Mountain Laboratories
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7204273280