Privacy-Preserving Medical Dialogue Generation Based on Federated Learning

Large-scale pre-trained dialogue models have shown outstanding performance across various dialogue-related natural language processing tasks. However, in privacy-sensitive domains like healthcare, concerns related to legal regulations and data security continue to pose challenges, resulting in data silos as a major barrier to building secure medical dialogue generation models. Federated learning is a distributed model training approach that allows models to be trained using data without the data leaving its local environment, making it an effective solution to address data silos in medial dialogue generation. In this paper, we focus on the task of medical dialogue generation, which utilizes medical dialogue data collected from three different Chinese short video platforms to train federated medical dialogue generation model. We employ the FedAvg algorithm to merge parameters of models trained on data from different sources. Experimental results demonstrate that in collaborative scenarios involving large organizations, federated learning effectively enhances the performance of medical dialogue models, improving the accuracy of output predictions. The effectiveness of federated learning varies among participants with different data volumes. Compared to the ideal scenario of centralized training, federated training yields an acceptable range of performance loss in the medical dialogue generation models.