FL-LoRaMAC: A Novel Framework for Enabling On-Device Learning for LoRa-Based IoT Applications

Abstract

The Internet of Things (IoT) enables us to gain access to a wide range of data from the physical world that can be analyzed for deriving critical state information. In this regard, machine learning (ML) is a valuable tool that can be used to develop models based on observed physical data, leading to efficient analytical decisions, including anomaly detection. In this work, we address some key challenges for applying ML in IoT applications that include maintaining privacy considerations of user data that are needed for developing ML models and minimizing the communication cost for transmitting the data over the IoT network. We consider a representative application of the anomaly detection of ECG signals that are obtained from a set of low-cost wearable sensors and transmitted to a central server using LoRaWAN, which is a popular and emerging low-power wide-area network (LPWAN) technology. We present a novel framework utilizing federated learning (FL) to preserve data privacy and appropriate features for uplink and downlink communications between the end devices and the gateway to optimize the communication cost. Performance results obtained from computer simulations demonstrate that the proposed framework leads to a 98% reduction in the volume of data that is required to achieve the same level of performance as in traditional centralized ML.