<p>Federated Learning (FL) is an innovative machine learning paradigm that enables collaborative model training among multiple participants without sharing raw data. Specifically, each participant trains a local model on its dataset and only needs to transmit the model updates or gradient information to build a global model. Although FL helps preserve data privacy, it still faces security concerns, particularly the threat posed by backdoor attacks. Backdoor attacks aim to insert hidden functionalities into the model, causing it to produce predetermined, incorrect outputs when triggered by specific inputs. Current research on backdoor attacks and defenses in FL settings primarily focuses on the image domain. With the growing application of FL in natural language processing, especially large language models, it is necessary to extend the research to these areas. Besides, current defense measures against backdoor attacks have several limitations: detection methods can become inefficient and less effective; mitigation techniques may have limited efficacy or require careful tuning; robust training may add computational overhead and resource requirements. These limitations are more pronounced in high-dimensional models. To counteract backdoor attacks and overcome such limitations, we propose MDSD, a multi-dimensional scaling-based defensive mechanism against backdoor attacks in federated learning. MDSD consists of three components: Clipping with Norm Threshold, Anomaly Detection, and Impact Control. Clipping with Norm Threshold limits the impact from potential malicious models by restricting each local model update under a specific threshold. Anomaly Detection employs the multi-dimensional scaling (MDS) technique to transform high-dimensional models into a lower-dimensional space while preserving the relative distances between data points. In the lower-dimensional space, the local model with the shortest distance to the global model is considered suspicious. Instead of directly discarding suspicious models, Impact Control lowers their impact by adjusting their aggregation weights. This can reduce the negative influence from mistakenly identifying a benign model as malicious. To conduct experiments, we utilized FedML, an open-source research library and benchmark for federated learning, to establish an experimental platform and simulate an FL environment. We then trained a transformer-based FL model for multi-classification tasks, during which we executed model-replacement-based backdoor attacks and applied MDSD. We also conducted experiments involving other defensive mechanisms for comparison. The experimental results indicate that MDSD has better overall defense performance. Specifically, MDSD can reduce the attack success rate from 100 to 0% with a slight decrease of less than 3% in model accuracy on main tasks. Our study not only underscores the novelty of applying MDS to detection techniques but also provides new insights for securing FL systems.</p>