<p>Federated Learning (FL) is a distributed machine learning framework that enables the efficient construction of a global model using a vast amount of decentralized client data. However, FL has little control over the local data and the training process, which is vulnerable to poisoning attacks. Among them, label flipping attacks significantly affect model robustness due to its low implementation threshold. Several approaches have been proposed to alleviate the risk of label flipping attacks, but their robustness is limited due to dependencies on various factors, such as a specific number of malicious clients or the availability of additional clean validation datasets. In this paper, we propose a method called FLGT, which guides server trust toward clients to robustly defend against label flipping attacks. It can obtain a trustworthy model update by utilizing Principal Component Analysis (PCA) to capture attack-related features from the similarity matrix of gradient angles in the last layer updates of local models. The central server first can assess the trustworthiness of clients to identify malicious ones by comparing the cosine similarity between the trustworthy model update and local model updates. Then, the central server can use the trustworthy model update to standardize the scale of local model updates, thereby enhancing system robustness. We conduct extensive evaluations of FLGT on CIFAR-10 and Fashion-MNIST, demonstrating its effectiveness against label flipping attacks and robustness in detecting malicious clients, particularly when the proportion of malicious clients ranges from 10% to 40%.</p>