Scholarlink

Federated learning (FL) is a technique that establishes a global model based on distributed data across multiple clients, while preserving data privacy. However, the frequent client-server communication in FL provides an opportunity for Byzantine clients, who can tamper with or falsify model parameters. Consequently, the presence of Byzantine clients impacts the convergence of the global model. To address this issue, we propose a framework that maintains Byzantine-robustness for FL in unreliable environments (BRFLUE). Specifically, BRFLUE consists of two components: a random-matching verification module and a credibility-table-based aggregation module. The former randomly matches clients and allows them to verify each other’s behavior anonymously. The latter updates clients’ credibility after receiving the client verification results and aggregates the global model based on the credibility table. The combination of these two modules can effectively identify possible Byzantine clients and reduce their adverse impact on the global model. Through implementations in different scenarios with classical datasets, we demonstrate the effectiveness of the proposed BRFLUE.

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

Byzantine-Robust Federated Learning for Unreliable Environments by Random-Matching Verification and Credibility Table

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yanjun Li,
  • Xiaomeng Li,
  • Hui Wan,
  • Cong Wang

摘要

Federated learning (FL) is a technique that establishes a global model based on distributed data across multiple clients, while preserving data privacy. However, the frequent client-server communication in FL provides an opportunity for Byzantine clients, who can tamper with or falsify model parameters. Consequently, the presence of Byzantine clients impacts the convergence of the global model. To address this issue, we propose a framework that maintains Byzantine-robustness for FL in unreliable environments (BRFLUE). Specifically, BRFLUE consists of two components: a random-matching verification module and a credibility-table-based aggregation module. The former randomly matches clients and allows them to verify each other’s behavior anonymously. The latter updates clients’ credibility after receiving the client verification results and aggregates the global model based on the credibility table. The combination of these two modules can effectively identify possible Byzantine clients and reduce their adverse impact on the global model. Through implementations in different scenarios with classical datasets, we demonstrate the effectiveness of the proposed BRFLUE.