Accelerated Federated Learning with Dynamic Model Partitioning for H-IoT

In the Healthcare Internet of Things (H-IoT), Federated Learning (FL) is a promising solution for processing huge amounts of medical data. At present, FL applied in H-IoT still faces many challenges such as low training efficiency and high data privacy risk. In this work, we develop a three-layer FL architecture, which introduces split learning to both prevent the leakage of medical data and improve training efficiency according to the inherent properties of Neural Networks (NN). Moreover, we formulate a long-term optimization problem with the goal of accelerating training speed of models in H-IoT. Then, an online model partitioning algorithm namely Privacy-aware Model Partitioning Algorithm (PMPA) is derived based on Lyapunov optimization theory that enables mobile devices of the FL architecture to efficiently train local models and protect the data privacy. Furthermore, the simulation results show that compared with traditional FL, the local training delay of the proposed algorithm can be reduced by 28.94% and 39.89%, respectively.