Scholarlink

In federated learning, data heterogeneity is a critical factor that constrains model performance. To overcome this challenge, we put forward a unique approach: FedCLF (Federated Contrastive Learning With Consensus Features). FedCLF integrates contrastive learning with federated learning based on consensus features, thus mitigating data heterogeneity through feature sharing and alignment mechanisms. Specifically, each client collaboratively trains a contrastive learning model and utilizes the trained encoder to extract features from local data, with a portion of these features being shared globally. During local model training, clients align the features extracted by their local models with those obtained in the previous stage to ensure model consistency and stability. Furthermore, after the model has been trained on their local datasets, clients employ the shared feature data to train the FC, thus further improving the model’s performance. It is shown by extensive experimental results that FedCLF significantly improves the accuracy and fairness of the model.

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

FedCLF: Contrastive Learning-Driven Framework for Mitigating Data Heterogeneity in Federated Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chunming Bai,
  • Heng Zhang,
  • Yong Xie,
  • Xin Su,
  • Xiangyuan Zhu,
  • Wenjie Kang,
  • Xuchong Liu

摘要

In federated learning, data heterogeneity is a critical factor that constrains model performance. To overcome this challenge, we put forward a unique approach: FedCLF (Federated Contrastive Learning With Consensus Features). FedCLF integrates contrastive learning with federated learning based on consensus features, thus mitigating data heterogeneity through feature sharing and alignment mechanisms. Specifically, each client collaboratively trains a contrastive learning model and utilizes the trained encoder to extract features from local data, with a portion of these features being shared globally. During local model training, clients align the features extracted by their local models with those obtained in the previous stage to ensure model consistency and stability. Furthermore, after the model has been trained on their local datasets, clients employ the shared feature data to train the FC, thus further improving the model’s performance. It is shown by extensive experimental results that FedCLF significantly improves the accuracy and fairness of the model.