Radio frequency identification and real-time security of LoRa devices with deep learning on embedded GPU

Abstract

In this paper, we propose the creation of unique Radio Frequency (RF) fingerprints by using communication signals for physical layer security of LoRa devices and real-time security control with these fingerprints. To achieve the goal, we constructed a 250 GB dataset, comprising 350,000 samples, by decoding raw LoRa communication data with Software Defined Radio and Embedded Linux and labeling each decoded ID with the corresponding raw LoRa signal. Utilizing the dataset, we conducted training in deep-learning models Convolutional Neural Networks, Temporal Convolutional Networks, Long Short-Term Memory, and Gated Recurrent Units, on a GPU server infrastructure. The CNN model demonstrated the best performance among the deep-learning models, achieving an accuracy rate of 99.8%. However, over time, we observed a decline in performance due to variations in the characteristics of electronic components. To ensure the stability of the system, we have determined that it is essential to retrain the model using the data acquired during secure communication. Low-power devices have limited packet transmissions and may not allocate a sufficient amount of data for retraining purposes. Hence, we performed minimum data analysis to retrain the system at regular intervals and found that even one packet per hour is sufficient for the devices. (c) 2017 Elsevier Inc. All rights reserved.