FedPLA: Cross-Silo Federated Learning with Proximal Regularization and Loss-Aware Aggregation

Cross-silo Federated Learning (FL) allows organizations to train models collaboratively while keeping data local, but performance often suffers under Non-IID data due to clients' diverse, domain-specific datasets. While existing work explores client-side personalization and server-side adaptivity separately, and some recent methods attempt to combine them, they often rely on complex optimization or static similarity measures, limiting their and robustness. We propose FedPLA, a cross-silo FL approach that explicitly integrates both perspectives. On the client side, FedPLA applies composite regularization to personalized heads, combining a proximal term for stability with L2 regularization to reduce overfitting. On the server side, a loss-aware aggregation adaptively adjusts global step sizes based on system loss dynamics, enhancing convergence under Non-IID data. Experiments on FMNIST, CIFAR-10, CIFAR-100, and a real-world power system dataset show that FedPLA consistently surpasses FL methods.