A federated learning and blockchain framework for physiological signal classification based on continual learning

Abstract

A key challenge of physiological signal processing in the Internet of Medical Things is that physiological signals usually have dynamic distribution changes. Another challenge is that patients' data is vulnerable to disclosures. To address these problems, we propose a meta continual learning method, called MetaCL. MetaCL consists of a shared feature extractor, a knowledge base, a micro classifier module and a blockchain module. The shared feature extractor adopts a horizontal federated learning mechanism to prevent data leakage. The knowledge base is built and updated based on a Splite-based method to overcome catastrophic forgetting. The micro classifier module uses a mean-based model transfer method to adapt to the emergence of new classes. It integrates a Kullback-Leibler divergence regularization to the loss function to deal with fuzzy boundaries of classes. The blockchain module is designed based on the Alliance chain to protect the privacy of the classification results. Experiment results show that refinement classification performance of MetaCL is 98.35%, which outperforms the compared state-of-the-art works. The backward transfer under four increment scenarios is all within -2.6%. At last, a blockchain-based engineering application is presented to show that MetaCL can prevent privacy protection in the Internet of Medical Things (IoMT).