FedSens: A Federated Learning Approach for Smart Health Sensing with Class Imbalance in Resource Constrained Edge Computing

Abstract

The advance of mobile sensing and edge computing has brought new opportunities for abnormal health detection (AHD) systems where edge devices such as smartphones and wearable sensors are used to collect people’s health information and provide early alerts for abnormal health conditions such as stroke and depression. The recent development of federated learning (FL) allows participants to collaboratively train powerful AHD models while keeping their health data private to local devices. This paper targets at addressing a critical challenge of adapting FL to train AHD models, where the participants’ health data is highly imbalanced and contains biased class distributions. Existing FL solutions fail to address the class imbalance issue due to the strict privacy requirements of participants as well as the heterogeneous resource constraints of their edge devices. In this work, we propose FedSens, a new FL framework dedicated to address the class imbalance problem in AHD applications with explicit considerations of participant privacy and device resource constraints. We evaluate FedSens using a real-world edge computing testbed on two real-world AHD applications. The results show that FedSens can significantly improve the accuracy of AHD models in the presence of severe class imbalance with low energy cost to the edge devices.