Improving Group Connectivity for Generalization of Federated Deep Learning