FedVG: Gradient-Guided Aggregation for Enhanced Federated Learning
Published in CVPR Findings 2026, 2026
This study introduces FedVG, a gradient-guided federated aggregation framework designed to improve learning under client data heterogeneity. FedVG uses a global validation set and layerwise validation gradient norms to adaptively weight client models based on their generalization behavior, rather than relying only on client dataset size. Experiments across natural and medical image benchmarks show that FedVG improves performance, particularly in highly heterogeneous federated learning settings.
Citation: Devkota, A., Thrasher, J., Adjeroh, D., Bhattarai, B. and Gyawali, P.K., 2026. FedVG: Gradient-Guided Aggregation for Enhanced Federated Learning. CVPR Findings.
Citation: Devkota, A., Thrasher, J., Adjeroh, D., Bhattarai, B. and Gyawali, P.K., 2026. FedVG: Gradient-Guided Aggregation for Enhanced Federated Learning. CVPR Findings.
Download Paper