Fair Federated Learning via Bounded Group Loss