On the Detectability of Active Gradient Inversion Attacks in Federated Learning