DeepLoRa: Fingerprinting LoRa Devices at Scale Through Deep Learning and Data Augmentation

Low-power wide-area networks (LPWANs) bring exceptional networking capabilities that will enable the massive roll-out of the Internet of Things (IoT). Among these capabilities are the support of ultra-low power consumption devices – up to 10 years of battery life – and connectivity up to tens of kilometers. The Long Range (LoRa) protocol has captured the research community’s attention due to its low cost, impressive sensitivity (better than -137dBm), and massive scalability potential. As tens of thousands of tiny LoRa devices are deployed over large geographic areas, a key component to the success of LoRa will be the development of reliable and robust authentication mechanisms. To this end, Radio Frequency FingerPrinting (RFFP) through deep learning (DL) has been heralded as an effective zero-power alternative to energy-hungry cryptography. Existing work on LoRa RFFP has mostly focused on small-scale testbeds and low-dimensional learning techniques; however, many challenges remain. Key among them are authentication techniques robust to a wide variety of channel variations and supporting a vast population of devices. In this work, we advance the state of the art by presenting the first massive experimental evaluation of DL RFFP and data augmentation techniques for LoRa. Specifically, we collect more than 1TB of waveform data from 100 devices over different deployment scenarios (outdoor vs. indoor) and spanning several days. We train and test several DL models (convolutional and recurrent neural networks) using either preamble or payload data slices. Finally, we propose a novel data augmentation technique called DeepLoRa to enhance LoRa RFFP performance. Results show that (i) the CNN models outperform the RNN-LSTM models in fingerprinting LoRa radios, (ii) using only payload data in the fingerprinting process outperforms preamble only data, (iii) DeepLoRa data augmentation technique improved classification accuracy from 19% to 36% in the problematic case of training on data collected at a different day than the testing data. Moreover, DeepLoR raises the RFFP accuracy from 82% to 91% when training and testing 100 bit-similar devices with data collected on the same day. As a significant additional contribution, sufficient models, code, and data used in this paper will be made available. Our dataset complies and extends the Signal Metadata Format (SigMF). We label the dataset using SigMF-metafiles. Moreover, we extend this format to incorporate a new extension designed to introduce LoRa dataset specifications.