A privacy-preserving decentralized randomized block-coordinate subgradient algorithm over time-varying networks

作     者：Wang, Lin Zhang, Mingchuan Zhu, Junlong Xing, Ling Wu, Qingtao 

作者机构：Henan Univ Sci & Technol Sch Informat Engn Luoyang 471023 Peoples R China 

出 版 物：《EXPERT SYSTEMS WITH APPLICATIONS》 (专家系统及其应用)

年 卷 期：2022年第208卷

核心收录：

学科分类：1201[管理学-管理科学与工程(可授管理学、工学学位)] 0808[工学-电气工程] 08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National Natural Science Foundation of China (NSFC) [62002102, 61971458, 61976243] Leading talents of science and technology in the Central Plain of China Scientific and Technological Innovation Talents of Colleges and Universities in Henan Province, China [22HASTIT014] 

主　　题：Convergence rate Privacy-preserving Randomized block-coordinate descent Subgradient projection 

摘      要：This study considers a constrained huge-scale optimization problem over networks where the objective is to minimize the sum of nonsmooth local loss functions. To solve this problem, many optimization algorithms have been proposed by employing (sub)gradient descent methods to handle high-dimensional data, but the computation of the entire sub(gradient) becomes a computational bottleneck. To reduce the computational burden of each agent and preserve privacy of data in time-varying networks, we propose a privacy-preserving decentralized randomized block-coordinate subgradient projection algorithm over time-varying networks, in which the coordinates of the subgradient vector is randomly chosen to update the optimized parameter and the partially homomorphic cryptography is used to protect the privacy of data. Further, we prove that our algorithm is convergent asymptotically. Moreover, the rates of convergence are also established by choosing appropriate step sizes, i.e., O(log K/K) under local strong convexity and O(log K/ K) under local convexity, in which K represents the number of iterations. Meanwhile, we show that the privacy of data can be protected by the proposed algorithm. The results of experiments demonstrate the computational benefit of our algorithm on two real-world datasets. The theoretical results are also verified by different experiments.