With the growing integration of large language models into distributed AI services, there is an increasing need to deploy them across decentralized environments while ensuring data privacy. Federated learning has emerged as a promising solution becaus...
With the growing integration of large language models into distributed AI services, there is an increasing need to deploy them across decentralized environments while ensuring data privacy. Federated learning has emerged as a promising solution because it enables collaborative training without directly collecting raw user data at a central server, but applying large language models in this setting remains challenging due to their large parameter size, high computational cost, and vulnerability to backdoor attacks. In this paper, these issues are investigated from two perspectives: the efficiency of LoRA-based federated learning and the robustness of representative defense mechanisms. The analysis shows that LoRA provides an efficient training framework, while perturbation-based and screening-based defenses exhibit different robustness characteristics and a clear trade-off between defense effectiveness and computational overhead. These results suggest that the practical deployment of large language models in federated learning requires a careful balance between efficiency and security, and that mechanistic interpretability-based defenses should be explored as an important direction for future work.