Efficient Federated Learning over Multiple Access Channel with Differential Privacy Constraints