Scholarlink

<p>Personalized federated learning is a method to solve the problem of non-independent and identically distributed data in the Internet of Things and provide personalized services to users. The extant personalized federated learning methods typically prioritize optimization at either the client or server. Our objective is to develop a mechanism that enables collaboration between these two parties. Therefore, we propose a personalized federated learning framework based on dual adaptive aggregation. Specifically, we design aggregation strategy based on dynamic similarity, which adaptively updates the cluster model on the server by exploiting the correlation between client and cluster distributions. Then, based on the differences in the impact of features on the feature-level aggregation results of local models, we construct feature-level local model aggregation strategy. Subsequently, Extensive experimentation was conducted on computer vision and Internet of Things benchmark datasets, and the results demonstrate that our method effectively enhances the accuracy of personalized local models.</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

Personalized federated learning based on dual adaptive aggregation in the IoT environments

  • 期刊论文

  • 发表日期：

    2025年8月30日

查看全文
  • Yanjin Cheng,
  • Wenmin Li,
  • Pengfei Gong,
  • Huimin Zhang,
  • Fei Gao,
  • Sujuan Qin,
  • Tengfei Tu

摘要

Personalized federated learning is a method to solve the problem of non-independent and identically distributed data in the Internet of Things and provide personalized services to users. The extant personalized federated learning methods typically prioritize optimization at either the client or server. Our objective is to develop a mechanism that enables collaboration between these two parties. Therefore, we propose a personalized federated learning framework based on dual adaptive aggregation. Specifically, we design aggregation strategy based on dynamic similarity, which adaptively updates the cluster model on the server by exploiting the correlation between client and cluster distributions. Then, based on the differences in the impact of features on the feature-level aggregation results of local models, we construct feature-level local model aggregation strategy. Subsequently, Extensive experimentation was conducted on computer vision and Internet of Things benchmark datasets, and the results demonstrate that our method effectively enhances the accuracy of personalized local models.