关键词: 联邦学习, 原型学习, 对比损失, 度量学习, 异质性数据处理

Abstract: To address issues such as significant model bias, unstable convergence, and poor generalization in federated learning due to non-independent and identically distributed (non-IID) client datasets and heterogeneous computing capabilities, this paper introduces a prototype-based model distance correction and aggregation algorithm for federated learning (FedMPD). FedMPD constructs local embedding networks on clients to extract features from heterogeneous data and corrects client models by introducing adjustments relative to both local and global prototypes. The algorithm incorporates prototype distance constraints, enabling clients to adaptively adjust training epochs based on the distance threshold between local and global prototypes, thereby mitigating the impact of device heterogeneity. During the model aggregation phase, FedMPD employs a weighted aggregation strategy that considers both the quantity of client data and the quality of local prototypes to more accurately quantify each client’s contribution to the global model. Experimental results demonstrate that FedMPD outperforms traditional federated learning algorithms in terms of model convergence stability, reduction in test loss, and improvement in test accuracy, providing a stable, efficient, and logically rigorous method for federated learning in heterogeneous environments.

Key words: federated learning, prototype learning, contrast loss, metric learning, heterogeneous data processing