Scholarlink

<p>Federated learning is a distributed machine learning method for training models with data localization. However, data generated by local devices is often heterogeneous (non-independent and identically distributed, Non-IID), as it is influenced by time, geographical location, and sampling bias, which can lead to suboptimal performance and slower convergence of models. Most methods mitigate this issue by adjusting the optimization direction on the local device or leveraging shared datasets. However, they still can’t work well for Non-IID data, due to the insufficient utilization of local high-quality data. For these problems, we propose an optimizing method, FedAEF. It trains autoencoders on local devices to extract local data features, and the server aggregates these features and generates independent and identically distributed (IID) data to fine-tune the global model. Experiments show that, compared to FedAVG, FedProx, and SCAFFOLD, FedAEF improves the accuracy by 2.5%, and the convergence performance of the model is improved by 55.9%.</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

FedAEF: optimizing federated learning with mining and enhancing local data features

  • 期刊论文

  • 发表日期：

    2025年8月30日

查看全文
  • Yan Zeng,
  • Chengchuang Huang,
  • Siyuan Teng,
  • Meiting Xue,
  • Yukun Shi,
  • Jilin Zhang,
  • Jian Wan,
  • Gangyong Jia

摘要

Federated learning is a distributed machine learning method for training models with data localization. However, data generated by local devices is often heterogeneous (non-independent and identically distributed, Non-IID), as it is influenced by time, geographical location, and sampling bias, which can lead to suboptimal performance and slower convergence of models. Most methods mitigate this issue by adjusting the optimization direction on the local device or leveraging shared datasets. However, they still can’t work well for Non-IID data, due to the insufficient utilization of local high-quality data. For these problems, we propose an optimizing method, FedAEF. It trains autoencoders on local devices to extract local data features, and the server aggregates these features and generates independent and identically distributed (IID) data to fine-tune the global model. Experiments show that, compared to FedAVG, FedProx, and SCAFFOLD, FedAEF improves the accuracy by 2.5%, and the convergence performance of the model is improved by 55.9%.