Privacy Masking Stochastic Subgradient-Push Algorithm for Distributed Online Optimization

被引:55
作者
Lu, Qingguo [1 ]
Liao, Xiaofeng [2 ]
Xiang, Tao [2 ]
Li, Huaqing [1 ]
Huang, Tingwen [3 ]
机构
[1] Southwest Univ, Coll Elect & Informat Engn, Chongqing Key Lab Nonlinear Circuits & Intelligen, Chongqing 400715, Peoples R China
[2] Chongqing Univ, Coll Comp, Chongqing 400044, Peoples R China
[3] Texas A&M Univ Qatar, Sci Program, Doha, Qatar
基金
中国国家自然科学基金;
关键词
Privacy; Cost function; Convex functions; Delays; Stochastic processes; Communication delays; differential privacy; distributed online optimization; stochastic subgradient-push algorithm; time-varying directed networks; MULTIAGENT NETWORKS; CONVEX-OPTIMIZATION; CONSENSUS;
D O I
10.1109/TCYB.2020.2973221
中图分类号
TP [自动化技术、计算机技术];
学科分类号
0812 ;
摘要
This article investigates the problem of distributed online optimization for a group of units communicating on time-varying unbalanced directed networks. The main target of the set of units is to cooperatively minimize the sum of all locally known convex cost functions (global cost function) while pursuing the privacy of their local cost functions being well masked. To address such optimization problems in a collaborative and distributed fashion, a differentially private-distributed stochastic subgradient-push algorithm, called DP-DSSP, is proposed, which ensures that units interact with in-neighbors and collectively optimize the global cost function. Unlike most of the existing distributed algorithms which do not consider privacy issues, DP-DSSP via differential privacy strategy successfully masks the privacy of participating units, which is more practical in applications involving sensitive messages, such as military affairs or medical treatment. An important feature of DP-DSSP is tackling distributed online optimization problems under the circumstance of time-varying unbalanced directed networks. Theoretical analysis indicates that DP-DSSP can effectively mask differential privacy as well as can achieve sublinear regrets. A compromise between the privacy levels and the accuracy of DP-DSSP is also revealed. Furthermore, DP-DSSP is capable of handling arbitrarily large but uniformly bounded delays in the communication links. Finally, simulation experiments confirm the practicability of DP-DSSP and the findings in this article.
引用
收藏
页码:3224 / 3237
页数:14
相关论文
共 47 条
[1]   Distributed Online Convex Optimization on Time-Varying Directed Graphs [J].
Akbari, Mohammad ;
Gharesifard, Bahman ;
Linder, Tamas .
IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, 2017, 4 (03) :417-428
[2]  
[Anonymous], 2017, ARXIV170705816
[3]  
[Anonymous], 2018, ARXIV180505573
[4]  
Cakan M, 2019, INT J MOL SCI
[5]   Distributed Algorithm Design for Nonsmooth Resource Allocation Problems [J].
Deng, Zhenhua ;
Nian, Xiaohong ;
Hu, Chen .
IEEE TRANSACTIONS ON CYBERNETICS, 2020, 50 (07) :3208-3217
[6]  
Durrett R, 2019, PROBABILITY, V49
[7]  
Dwork C, 2006, LECT NOTES COMPUT SC, V4052, P1
[8]   Differentially Private Consensus With an Event-Triggered Mechanism [J].
Gao, Lan ;
Deng, Shaojiang ;
Ren, Wei .
IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, 2019, 6 (01) :60-71
[9]  
Ge Rong, 2015, P 28 C LEARNING THEO
[10]   RECURSIVE STOCHASTIC ALGORITHMS FOR GLOBAL OPTIMIZATION IN RD [J].
GELFAND, SB ;
MITTER, SK .
SIAM JOURNAL ON CONTROL AND OPTIMIZATION, 1991, 29 (05) :999-1018