Scholarlink

<p>Federated learning (FL) has becoming a prevailing paradigm which enables small-scale devices to collaboratively learn a shared model efficiently and trains a machine learning model without exchanging data. However, though the original data never leave the local machines in federated learning, possible privacy leakage still exists. To make strong privacy guarantee, in this paper, we incorporate the notion of differential privacy (DP) to study the federated averaging (<i>FedAvg</i>) algorithm. In particular, by adding calibrated gaussian noise, we propose a set of differentially private federated averaging algorithms (<i>DP-FedAvg</i>) under the full and partial participation schemes. We provide tight analysis of the privacy bound by using advanced composition and privacy amplification techniques. We also analyze the convergence bound of <i>DP-FedAvg</i> without having the assumptions: (i) the data are the independent identically distribution (IID), and (ii) all the devices are active. It turns out that the convergence rate is consistent with the one without DP guarantee. The effectiveness of our algorithms is demonstrated by synthetic and real datasets.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Privacy-Preserving Federated Averaging on Heterogeneous Data

  • 期刊论文

  • 发表日期：

    2025年10月15日

查看全文
  • Lingjie Zhang,
  • Xiao Guo,
  • Hai Zhang

摘要

Federated learning (FL) has becoming a prevailing paradigm which enables small-scale devices to collaboratively learn a shared model efficiently and trains a machine learning model without exchanging data. However, though the original data never leave the local machines in federated learning, possible privacy leakage still exists. To make strong privacy guarantee, in this paper, we incorporate the notion of differential privacy (DP) to study the federated averaging (FedAvg) algorithm. In particular, by adding calibrated gaussian noise, we propose a set of differentially private federated averaging algorithms (DP-FedAvg) under the full and partial participation schemes. We provide tight analysis of the privacy bound by using advanced composition and privacy amplification techniques. We also analyze the convergence bound of DP-FedAvg without having the assumptions: (i) the data are the independent identically distribution (IID), and (ii) all the devices are active. It turns out that the convergence rate is consistent with the one without DP guarantee. The effectiveness of our algorithms is demonstrated by synthetic and real datasets.