How to Defend Against Large-scale Model Poisoning Attacks in Federated Learning: A Vertical Solution