Do Gradient Inversion Attacks Make Federated Learning Unsafe?