FedMABA: Towards Fair Federated Learning through Multi-Armed Bandits Allocation