LORA PROTOCOL FOR DECENTRALIZED FEDERATED LEARNING ON RESOURCE-CONSTRAINED DEVICES

Abstract

Internet of Things (IoT) is an ever-growing field of computer science. Current trends require the IoT nodes to solve more complicated problems, such as classification and prediction. This is achieved by utilizing the power of Artificial Intelligence (AI), such as running Machine Learning (ML) and Neural Networks (NNs) algorithms on the edge. Federated Learning (FL) is an AI paradigm that allows devices to make the global model without sharing local data, thus preserving privacy. However, the regular FL algorithms are based on centralized aggregators that control the process of creating the global model. Decentralized FL (DFL) is an approach where every node communicates with the neighbouring nodes to create the global model. This work presents an efficient LoRa protocol to facilitate DFL on resource-constrained devices. Feasibility tests’ results show up to 30.4% increase in the accuracy after 10 DFL rounds on the Heltec WiFi LoRa V3 nodes for the MLP MNIST model. The protocol allows for the parallel model transfer by employing the multi-channel transmission paradigm. Simulation results show that the multi-channel model transmission is up to 10% more effective than transmission in a single channel.