Class aware efficient client selection and multi-loss guided local performance optimization in federated learning

Federated Learning (FL) enables decentralized model training across multiple edge devices, preserving privacy through local data processing at the edge. While FL reduces data transfer by exchanging only model weights, inherent multiple aggregation steps can increase the communication overhead. Practically, clients’ datasets exhibit significant variation in their label distributions. This means that data across clients is not independently and identically distributed (non-IID) with respect to label classes, creating distinct patterns unique to each client. It causes inconsistencies between local objectives and global optima, which impacts the overall performance of global training. Additionally, trade-offs between performance and energy consumption remain a key challenge mostly ignored by existing FL methods. To address this issue, we introduce a class distribution-aware client selection algorithm guided by a multi-loss function that optimizes both performance and energy consumption across participating edge devices, named CACS-FL. In CACS-FL, firstly, a confidence score is calculated for each edge device based on their heterogeneity score and energy consumption, which is further optimized to select an appropriate set of edge devices. After selection, the clients perform local training using a weighted multi-loss function, which improves personalized performance and achieves higher global performance. Experimental demonstrations on various datasets showcase CACS-FL’s advantages over existing state-of-the-art approaches. CACS-FL also guarantees faster convergence with fairness in performance across the participating edge devices.