Scholarlink

<p>Federated learning allows clients to collaboratively train models while keeping client data local. Initially, it trains a global model to serve all clients, but when the distribution of each local dataset differs significantly from the global dataset, the local objectives of each client may diverge from the globally optimal values, leading to drift in local updates. This phenomenon greatly impacts model performance. The primary purpose of client participation in federated learning is to obtain personalized models with better local performance. In order to solve this problem, this paper proposes a new federated learning algorithm - Federated Learning with Adaptive Intra - cluster Aggregation (FedACC). This algorithm utilizes the inference correlation of client-uploaded models to the server and divides clients with similar data distributions into clusters. During the weighted aggregation of models within each cluster, we introduce an adaptive weight learning algorithm and use the obtained weights to perform the weighted aggregation of cluster models. The algorithm can cluster clients with similar data distributions and utilize adaptive weight learning within the cluster to obtain optimal aggregation weights, enabling more efficient and personalized federated learning through a weighted aggregation cluster model. Our experiments are conducted on three public image datasets, namely MNIST, Fashion - MNIST, and CIFAR − 10, and in a data - heterogeneous environment. Compared with three baseline algorithms, the Federated Averaging algorithm (FedAvg), the Federated Proximal algorithm (FedProx), and the Federated Learning with Intra - cluster Similarity algorithm (FLIS), the global model of the FedACC algorithm proposed in this paper converges faster and has a higher accuracy. On the Fashion - MNIST dataset, compared with FedAvg, FedProx, and FLIS algorithms, the accuracy of FedACC is improved by 11.6%, 10.5%, and 3.0% respectively, which proves the effectiveness of the FedACC algorithm.</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 with adaptive aggregation within clusters

  • 期刊论文

  • 发表日期：

    2025年8月29日

查看全文
  • Shiyuan Ding,
  • Yanhong Liu,
  • Haobin Shi,
  • Yingying Gao

摘要

Federated learning allows clients to collaboratively train models while keeping client data local. Initially, it trains a global model to serve all clients, but when the distribution of each local dataset differs significantly from the global dataset, the local objectives of each client may diverge from the globally optimal values, leading to drift in local updates. This phenomenon greatly impacts model performance. The primary purpose of client participation in federated learning is to obtain personalized models with better local performance. In order to solve this problem, this paper proposes a new federated learning algorithm - Federated Learning with Adaptive Intra - cluster Aggregation (FedACC). This algorithm utilizes the inference correlation of client-uploaded models to the server and divides clients with similar data distributions into clusters. During the weighted aggregation of models within each cluster, we introduce an adaptive weight learning algorithm and use the obtained weights to perform the weighted aggregation of cluster models. The algorithm can cluster clients with similar data distributions and utilize adaptive weight learning within the cluster to obtain optimal aggregation weights, enabling more efficient and personalized federated learning through a weighted aggregation cluster model. Our experiments are conducted on three public image datasets, namely MNIST, Fashion - MNIST, and CIFAR − 10, and in a data - heterogeneous environment. Compared with three baseline algorithms, the Federated Averaging algorithm (FedAvg), the Federated Proximal algorithm (FedProx), and the Federated Learning with Intra - cluster Similarity algorithm (FLIS), the global model of the FedACC algorithm proposed in this paper converges faster and has a higher accuracy. On the Fashion - MNIST dataset, compared with FedAvg, FedProx, and FLIS algorithms, the accuracy of FedACC is improved by 11.6%, 10.5%, and 3.0% respectively, which proves the effectiveness of the FedACC algorithm.