A federated binarized neural network model for constrained devices in IoT healthcare services

Abstract

In IoT healthcare environment, the devices are not sufficiently powerful for operating recent deep learning models, and data collected by the devices are usually decentralized. Moreover, data are unavailable to share between devices because of information security issues. Therefore, a concept of federated learning has emerged to overcome data sharing issues, and a concept of binarized neural network has emerged to generate lightweight deep learning models. This paper proposes a federated binarized neural network model to derive a reliable healthcare system in this circumstance. This paper shows an overview of considered system model with constrained IoT healthcare devices. In addition, this paper shows illustrations of implementing the proposed federated learning model with the proposed binarized MLP networks by utilizing an open-source library. The experiment results show that the binarized MLP network shows comparable performances compared to the full-precision MLP network while the binarized MLP requires about 10-times less model size for training.