Towards Fast Network Intrusion Detection based on Efficiency-preserving Federated Learning