Scholarlink

<p>With the growth of the Internet of Things (IoT) and communication technologies, edge devices have become more diverse. This diversity has increased the computational load on these systems and led to differences between devices. In mobile edge computing, variations in communication and computing resources can prevent some devices from updating models quickly. This delay affects overall performance. In addition, in federated learning, data that is not independently and identically distributed (non-IID) makes it hard for clients to maintain personalized models.To address these issues, this paper introduces a personalized federated learning framework. This framework enhances the resource allocation optimization algorithm by dynamically adjusting the depth of model inference and the bandwidth allocation strategy, which assists devices with limited computational capabilities in completing inference tasks promptly. Furthermore, it divide the client models into global and personalized layers. Only the global layers are combined, which helps manage the diversity in data distributions. Simulation results show that the proposed FedMEM method is superior to other state-of-the-art methods, and can drastically reduce system latency.</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

FedMEM: Adaptive Personalized Federated Learning Framework for Heterogeneous Mobile Edge Environments

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年4月14日

查看全文
  • Chen Ximing,
  • He Xilong,
  • Cheng Du,
  • Wu Tiejun,
  • Tian Qingyu,
  • Chen Rongrong,
  • Qiu Jing

摘要

With the growth of the Internet of Things (IoT) and communication technologies, edge devices have become more diverse. This diversity has increased the computational load on these systems and led to differences between devices. In mobile edge computing, variations in communication and computing resources can prevent some devices from updating models quickly. This delay affects overall performance. In addition, in federated learning, data that is not independently and identically distributed (non-IID) makes it hard for clients to maintain personalized models.To address these issues, this paper introduces a personalized federated learning framework. This framework enhances the resource allocation optimization algorithm by dynamically adjusting the depth of model inference and the bandwidth allocation strategy, which assists devices with limited computational capabilities in completing inference tasks promptly. Furthermore, it divide the client models into global and personalized layers. Only the global layers are combined, which helps manage the diversity in data distributions. Simulation results show that the proposed FedMEM method is superior to other state-of-the-art methods, and can drastically reduce system latency.