On the Efficiency of Privacy Attacks in Federated Learning