FedHPro: Federated Hyper-Prototype Learning via Gradient Matching

Researchers introduced FedHPro, a federated learning method that learns hyper-prototypes via gradient matching instead of averaging local prototypes.

The approach is designed to reduce semantic drift, strengthen inter-class separation, and improve intra-class consistency under non-IID data.

By better aligning global class representations across clients, FedHPro delivers stronger generalization and state-of-the-art results on multiple benchmarks.