【摘要】 为了解决传统联邦蒸馏中由数据异质造成的知识误导现象，针对医疗对话文本生成这种复杂任务，提出了一种通过动态知识融合和客户端选择进行蒸馏的模型FedKS,对知识的积累和传递进行更加精细的建模.首先，在联邦学习中设计了一种有效的知识聚合机制.其次，针对存在误导知识这一问题，提出了一种基于阈值的方法来优化每个客户端的本地模型更新选择.通过计算各客户端对全局模型性能的增益，决定是否采用知识蒸馏后的本地模型.FedKS模型可以有效解决由于误导知识导致的局部模型性能下降问题，从而实现高效的知识聚合和传递.基于多个数据集的实验表明，FedKS模型比现有基线模型训练收敛速度更快，性能更优异，并能够支持异质模型.更多还原

【Abstract】 To address the knowledge misleading problem caused by data heterogeneity in traditional federated distillation, a model for generating medical dialogue texts through dynamic knowledge fusion and clients selection for distillation, FedKS, is proposed, providing a more detailed modeling for knowledge accumulation and transmission procedure. Firstly, an effective knowledge aggregation mechanism is designed in federated learning. Secondly, to address the issue of misleading knowledge, a threshold-based method is proposed to optimize the updating selection of local model on each client. By calculating the performance gain of the global model on each client, it is determined whether to adopt the local model after knowledge distillation. FedKS model can effectively address the problem of local model performance degradation caused by misleading knowledge, thereby achieving efficient knowledge aggregation and transmission. Experiments based on multiple benchmark datasets show that FedKS model accelerates training convergence speed and improves the performance compared with the existing baselines. In addition, it can support heterogeneous client models.更多还原