Scholarlink

<p>In recent federated learning (FL) research, while significant progress has been made in addressing data and model heterogeneity, the challenges posed by computational heterogeneity in devices remain largely unresolved. In this work, we propose PHFL, a novel FL framework that leverages a hybrid synchronization mechanism to tackle the computational heterogeneity challenges arising from heterogeneous device environments. PHFL introduces a unique combination of intragroup synchronous updates and intergroup asynchronous communication, ensuring efficient aggregation and improved model performance across diverse devices. Unlike traditional FL methods, which rely on direct parameter sharing, PHFL transmits feature representations <i>C</i>, facilitating aggregation at the server level and supporting personalized local updates through a weighted combination of global aggregated features and local loss functions. To validate the effectiveness of PHFL, we conduct extensive experiments comparing it against prominent FL algorithms, including but not limited to FedAvg, FedProto, PerAvg, and FedBuff, across multiple datasets (CIFAR10, MNIST, and EMNIST) and model architectures (CNN and ResNet). Our results demonstrate that PHFL outperforms the other methods in terms of accuracy and convergence speed, particularly under non-IID data distributions. This work highlights the potential of hybrid synchronization strategies and temporal aggregation in enhancing FL frameworks for real-world heterogeneous environments.</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

PHFL: a federated learning framework based on a hybrid mechanism

  • 期刊论文

  • 发表日期：

    2025年3月16日

查看全文
  • Xingyu Chen,
  • Yuxiang Chen,
  • Wei Liang,
  • Dacheng He,
  • Kuanching Li,
  • Mirjana Ivanović

摘要

In recent federated learning (FL) research, while significant progress has been made in addressing data and model heterogeneity, the challenges posed by computational heterogeneity in devices remain largely unresolved. In this work, we propose PHFL, a novel FL framework that leverages a hybrid synchronization mechanism to tackle the computational heterogeneity challenges arising from heterogeneous device environments. PHFL introduces a unique combination of intragroup synchronous updates and intergroup asynchronous communication, ensuring efficient aggregation and improved model performance across diverse devices. Unlike traditional FL methods, which rely on direct parameter sharing, PHFL transmits feature representations C, facilitating aggregation at the server level and supporting personalized local updates through a weighted combination of global aggregated features and local loss functions. To validate the effectiveness of PHFL, we conduct extensive experiments comparing it against prominent FL algorithms, including but not limited to FedAvg, FedProto, PerAvg, and FedBuff, across multiple datasets (CIFAR10, MNIST, and EMNIST) and model architectures (CNN and ResNet). Our results demonstrate that PHFL outperforms the other methods in terms of accuracy and convergence speed, particularly under non-IID data distributions. This work highlights the potential of hybrid synchronization strategies and temporal aggregation in enhancing FL frameworks for real-world heterogeneous environments.