Developing Machine Learning Techniques to Identify LoRa Devices Using the Radio Frequency Fingerprint

Abstract

This thesis investigates IoT device authentication through LoRa device classification. LoRa devices are sensors that send small packets of information, such as temperature information. We introduce a scheme for authenticating LoRa devices within a network to avoid misleading information from an intruder device. Traditionally, deep learning successfully classifies LoRa devices using Convolution Neural Networks (CNNs). In prior research, Radio Frequency Fingerprint Identification (RFFI) is the ideal classification method because it does not affect the IoT device's efficiency. RFFI is a lightweight device authentication scheme that relies on conventional signal processing that exposes the unique hardware impairments that are device-specific. One of the gaps in the LoRa-RFFI research is the training data that consists of a single Spreading Factor (SF). This thesis addresses this gap by providing a LoRa-RFFI study using a dataset with more than one SF. We propose a hybrid RFFI scheme combining deep learning benefits in feature extraction and machine learning in low-cost classification. Then we investigate a less complex method for an IQ sample transformation not covered in the literature on LoRa-RFFI. We investigate five deep models for device classification, and feature extraction, including the SwinTransformer (ST) and the Few Shots Learning (FSL) models not covered in the literature. The proposed LoRa-RFFI system is successful at classifying LoRa devices with small costs.

Date of Award	2023
Original language	American English
Awarding Institution	HBKU College of Science and Engineering