Automatic Generation of CNN Models For Radiofrequency Fingerprinting

This paper explores the automatic generation of convolutional neural network (CNN) models for radiofrequency (RF) fingerprinting. Device fingerprinting aims to identify a wireless transmitter based on the unique waveform it emits due to minor hardware differences introduced during manufacturing. Our proposal is to use Gen-CNN, a novel hybrid methodology that uses genetic algorithms to automatically generate optimized, reduced CNNs for classification. Originally developed for medical imaging, we extend the application of Gen-CNN to RF fingerprinting of long-range (LoRa) devices, which are low-power transmitters designed for Internet of Things (IoT) connectivity. The datasets used in this work are created with spectrograms derived from RF data captured with a software-defined radio (SDR). This paper offers a novel application of automated machine learning algorithms for classifying spectrograms aimed at enhancing security protocols in wireless networks, something that, to the best of our knowledge, has never been attempted before.