SplitGP: Achieving Both Generalization and Personalization in Federated Learning