Scholarlink

Asynchronous federated learning (AFL) mitigates scalability and straggler issues in federated training but exacerbates communication bottlenecks due to frequent high-dimensional gradient transmissions. While gradient quantization techniques reduce communication overhead, existing static compression strategies fail to adapt to the dynamic nature of training phases, leading to suboptimal trade-offs between convergence performance and communication budget. This paper proposes FedBuff_AGQ, an asynchronous FL framework that integrates adaptive gradient quantization with buffered aggregation to enhance communication efficiency while preserving convergence guarantees. We theoretically analyze the impact of quantization errors and develop an adaptive quantization strategy using Recursive Least Squares (RLS) estimation and rolling-horizon optimization to balance convergence efficiency and communication constraints. Experiments on benchmark FL datasets demonstrate that FedBuff_AGQ achieves superior accuracy under the same communication budget, highlighting its effectiveness in improving both communication efficiency and model performance.

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

Adaptive Gradient Quantization for Efficient Federated Learning with Buffered Asynchronous Aggregation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Wei Fan,
  • Minhan Hu,
  • Keke Yang,
  • Jing Li

摘要

Asynchronous federated learning (AFL) mitigates scalability and straggler issues in federated training but exacerbates communication bottlenecks due to frequent high-dimensional gradient transmissions. While gradient quantization techniques reduce communication overhead, existing static compression strategies fail to adapt to the dynamic nature of training phases, leading to suboptimal trade-offs between convergence performance and communication budget. This paper proposes FedBuff_AGQ, an asynchronous FL framework that integrates adaptive gradient quantization with buffered aggregation to enhance communication efficiency while preserving convergence guarantees. We theoretically analyze the impact of quantization errors and develop an adaptive quantization strategy using Recursive Least Squares (RLS) estimation and rolling-horizon optimization to balance convergence efficiency and communication constraints. Experiments on benchmark FL datasets demonstrate that FedBuff_AGQ achieves superior accuracy under the same communication budget, highlighting its effectiveness in improving both communication efficiency and model performance.