FedSPC: Shared Parameter Correction for Personalized Federated Learning
Researchers have introduced FedSPC, a novel method designed to enhance personalized federated learning (PFL). This approach specifically targets the shared parameters within PFL models, applying a control-variate correction to address inconsistencies that arise from clients optimizing different local objectives. FedSPC can be integrated into various PFL architectures and has demonstrated performance improvements across several common PFL methods on benchmark datasets like CIFAR-100 and Tiny-ImageNet, utilizing models such as ViT, ResNet-34, and VGG-11. AI