Scholarlink

<p>Fast clients update the global model frequently in semi-asynchronous federated learning (Semi-AFL) frameworks. As a result, the global model fits well fast clients’ local data but learns little knowledge from slow clients. The update imbalance significantly degrades the performance of Semi-AFL frameworks. Especially under heavy-tailed distributions of clients’ response time, it takes several tens of rounds for a large proportion of slow clients to update the global model. We propose a novel Semi-AFL framework called FederatedBalancing (or FedBal) against the imbalance of client updates. It is aware of the response time distribution of clients and alleviates the increasing imbalance with iteration rounds via an online self-adaptive grouping mechanism that dynamically assigns clients to fast or slow groups and improves the proportion of slow clients participating in model aggregations. We evaluate FedBal and the state-of-the-art Semi-AFL frameworks on four open datasets under three typical distributions of clients’ response time. Experiment results show that FedBal significantly improves the proportion of slow clients updating the global model and achieves higher accuracy and more balanced updates.</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

FedBal: a self-adaptive and semi-asynchronous federated learning framework against the imbalance of client updates

  • 期刊论文

  • 发表日期：

    2026年7月8日

查看全文
  • Zhiyao Hu,
  • Junjie Xie

摘要

Fast clients update the global model frequently in semi-asynchronous federated learning (Semi-AFL) frameworks. As a result, the global model fits well fast clients’ local data but learns little knowledge from slow clients. The update imbalance significantly degrades the performance of Semi-AFL frameworks. Especially under heavy-tailed distributions of clients’ response time, it takes several tens of rounds for a large proportion of slow clients to update the global model. We propose a novel Semi-AFL framework called FederatedBalancing (or FedBal) against the imbalance of client updates. It is aware of the response time distribution of clients and alleviates the increasing imbalance with iteration rounds via an online self-adaptive grouping mechanism that dynamically assigns clients to fast or slow groups and improves the proportion of slow clients participating in model aggregations. We evaluate FedBal and the state-of-the-art Semi-AFL frameworks on four open datasets under three typical distributions of clients’ response time. Experiment results show that FedBal significantly improves the proportion of slow clients updating the global model and achieves higher accuracy and more balanced updates.