Weighted distributed differential privacy ERM Convex and non-convex

被引:2
作者
Kang, Yilin [1 ,4 ]
Liu, Yong [1 ,2 ,3 ]
Niu, Ben [1 ]
Wang, Weiping [1 ]
机构
[1] Chinese Acad Sci, Inst Informat Engn, 89-A Minzhuang Rd, Beijing 100093, Peoples R China
[2] Renmin Univ China, Gaoling Sch Artificial Intelligence, 59 Zhongguancun St, Beijing 100872, Peoples R China
[3] Beijing Key Lab Big Data Management & Anal Method, Beijing, Peoples R China
[4] Univ Chinese Acad Sci, Sch Cyber Secur, Beijing, Peoples R China
基金
中国国家自然科学基金;
关键词
Distributed machine learning; Differential privacy; Weighted parties; Empirical risk minimization; Strongly convex; Polyak-Ł ojasiewicz condition;
D O I
10.1016/j.cose.2021.102275
中图分类号
TP [自动化技术、计算机技术];
学科分类号
0812 ;
摘要
Distributed machine learning allows different parties to learn a single model over all data sets without disclosing their own data. In this paper, we propose a weighted distributed differentially private (WD-DP) empirical risk minimization (ERM) method to train a model in distributed setting, considering different weights of different clients. For the first time, we theoretically analyze the benefits brought by weighted paradigm in distributed differentially private machine learning. Our method advances the state-of-the-art differentially private ERM methods in distributed setting. By detailed theoretical analysis, we show that in distributed setting, the noise bound and the excess empirical risk bound can be improved by considering different weights held by multiple parties. Additionally, in some situations, the constraint: strongly convexity of the loss function in ERM is not easy to achieve, so we generalize our method to the condition that the loss function is not restricted to be strongly convex but satisfies the Polyak-Lojasiewicz condition. Experiments on real data sets show that our method is more reliable and we improve the performance of distributed differentially private ERM, especially in the case that data scales on different clients are uneven. Moreover, it is an attractive result that our distributed method achieves almost the same theoretical and experimental results as previous centralized methods. (c) 2021 Elsevier Ltd. All rights reserved.
引用
收藏
页数:18
相关论文
共 61 条
  • [1] Deep Learning with Differential Privacy
    Abadi, Martin
    Chu, Andy
    Goodfellow, Ian
    McMahan, H. Brendan
    Mironov, Ilya
    Talwar, Kunal
    Zhang, Li
    [J]. CCS'16: PROCEEDINGS OF THE 2016 ACM SIGSAC CONFERENCE ON COMPUTER AND COMMUNICATIONS SECURITY, 2016, : 308 - 318
  • [2] [Anonymous], 2014, P INT C MACH LEARN
  • [3] Arora Raman, 2019, NEURIPS, P13378
  • [4] Backes M., 2016, P 2016 ACM SIGSAC C, P319, DOI [10.1145/, 10.1145/2976749.2978355]
  • [5] Private Empirical Risk Minimization: Efficient Algorithms and Tight Error Bounds
    Bassily, Raef
    Smith, Adam
    Thakurta, Abhradeep
    [J]. 2014 55TH ANNUAL IEEE SYMPOSIUM ON FOUNDATIONS OF COMPUTER SCIENCE (FOCS 2014), 2014, : 464 - 473
  • [6] Bernstein G., 2019, ADV NEURAL INFORM PR, P525
  • [7] Bontempi Gianluca, 2018, ULB The Machine Learning Group
  • [8] Distributed optimization and statistical learning via the alternating direction method of multipliers
    Boyd S.
    Parikh N.
    Chu E.
    Peleato B.
    Eckstein J.
    [J]. Foundations and Trends in Machine Learning, 2010, 3 (01): : 1 - 122
  • [9] Concentrated Differential Privacy: Simplifications, Extensions, and Lower Bounds
    Bun, Mark
    Steinke, Thomas
    [J]. THEORY OF CRYPTOGRAPHY, TCC 2016-B, PT I, 2016, 9985 : 635 - 658
  • [10] Carlini N, 2019, PROCEEDINGS OF THE 28TH USENIX SECURITY SYMPOSIUM, P267