Federated Learning with Differential Privacy: An Utility-Enhanced Approach