Efficient Differentially Private Secure Aggregation for Federated Learning via Hardness of Learning with Errors.