Low Complexity Deep Learning Models for LoRa Radio Frequency Fingerprinting

LoRaWAN devices are secured using traditional cryptographic methods. However, the end devices are still vulnerable to security attacks such as impersonation. To counter these attacks, LoRa requires an additional layer of security at the physical level. Deep Learning-based LoRa device identification using Radio Frequency Fingerprinting is currently seen as a key candidate for enhancing LoRa security at the physical layer. However, a more in-depth study of certain aspects of this approach is required.
Firstly, the appropriate LoRa signal representation has to be chosen which provides high device identification accuracy. Secondly, there is a need for deep learning models with less computational complexity and high performance. This paper contributes to the state-of-the-art in two ways: (1) we evaluate various signal representations such as raw In-phase Quadrature (IQ), Amplitude and Phase ($A/\phi$), frequency domain (FFT), and time-frequency domain (Spectrogram) (2) we show that an existing complex ResNet model can be optimized into a lightweight model by tuning its parameters without compromising much on the performance. We implement a different lightweight model for IQ, FFT, and $A/\phi$ representations. We show that our optimized ResNet model for spectrogram and proposed lightweight model for sequential data can achieve an accuracy of over 97\%.