Differentially Private Federated Learning with Local Regularization and Sparsification