Prevention over Correction: Learning Aligned Representations in One-shot Federated Learning