Scholarlink

Federated learning (FL) has emerged as a leading approach for decentralized model training, preserving data privacy by exchanging only model parameters. However, recent studies have exposed vulnerabilities, revealing that attackers can reconstruct private data from shared model parameters, compromising user privacy. To address this, various defense methods have been proposed, with gradient perturbation being a lightweight yet effective strategy. However, it often incurs a trade-off with model performance. In this paper, we introduce a lightweight defense mechanism tailored for Federated Stochastic Gradient Descent (FedSGD). Our method leverages maximum gradient values to gauge noise scale, while also preserving and compensating for previous noise iterations to mitigate performance impacts.

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

Using Local Gradient Compensation in Federated Learning for Improving Privacy and Performance

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jia-Wei Chang,
  • Wei-Cheng Chen,
  • Chih-Chieh Hung

摘要

Federated learning (FL) has emerged as a leading approach for decentralized model training, preserving data privacy by exchanging only model parameters. However, recent studies have exposed vulnerabilities, revealing that attackers can reconstruct private data from shared model parameters, compromising user privacy. To address this, various defense methods have been proposed, with gradient perturbation being a lightweight yet effective strategy. However, it often incurs a trade-off with model performance. In this paper, we introduce a lightweight defense mechanism tailored for Federated Stochastic Gradient Descent (FedSGD). Our method leverages maximum gradient values to gauge noise scale, while also preserving and compensating for previous noise iterations to mitigate performance impacts.