A Federated Adversarial Learning Method for Biomedical Named Entity Recognition

Identifying medical terms with specific meaning and information with semantic attribute is the prerequisite of conducting semantic analysis in medical field. However, the problem of medical data island restricts the development of entity recognition in a great extent. In addition to the ban of data sharing between different hospitals, different departments in the same hospital also can not exchange data due to privacy security concerns and ethical issues. To solve these problems, in the federated learning framework, the server trains a global model collaboratively through aggregating the encrypted or noised model parameters of the local participated clients without data leakage. In this paper, to well apply federated learning on biomedical named entity recognition (BioNER), we propose the federated adversarial learning (FAL) method with consideration of the training cost and model performance. FAL not only makes use of a modified structured pruning scheme to reduce the number of model parameters but also exploits an improved adversarial learning approach named protected fast gradient method (PFGM) to enhance the robustness and generalization of the model. In the experiment, we use the datasets of five departments in the same tumor hospital, such as gynecology department and gastric surgery department. Results show that the proposed FAL framework achieves expected effect with high efficiency.