RF Fingerprinting Based on Reservoir Computing Using Narrowband Optoelectronic Oscillators

Radiofrequency (RF) fingerprinting refers to a range of technologies that recognize transmitters by their intrinsic hardware-level characteristics. These characteristics are often introduced during the fabrication process and form a unique fingerprint of the transmitter that is very hard to counterfeit. RF fingerprinting often serves as a security measure at the physical-layer of communication networks against potentials attacks. In recent years, neuromorphic computing techniques such as convolutional neural networks (CNNs) have been explored as classifiers for RF fingerprinting. However, in radiofrequency communication networks, the transmitted signals are I/Q modulated on multi-GHz carriers while most conventional machine learning algorithms operate at the baseband. Therefore, the I/Q modulated signals have to be demodulated and converted into compatible formats before applying to these platforms – a procedure that inevitably slows down the processing speed. Moreover, the deep learning technologies often require a large amount of data to train the artificial neural networks (ANNs) while in practice, the available amount of data for a new transmitter is limited. Reservoir computing (RC) provides a relatively simple yet powerful structure that is capable of reaching state-of-the-art performance on several benchmarks. However, traditional digital RC also operates at baseband, which is not suitable for directly processing the I/Q modulated signals. In this article, we propose a reservoir computer based on narrowband optoelectronic oscillator (OEO) that can be utilized to directly classify I/Q modulated signals without the need for demodulation. We successfully train and test our narrowband OEO-based RC on three publicly available benchmarks, namely the FIT/CorteXlab RF fingerprinting dataset, the ORACLE RF fingerprinting dataset, and the AirID RF fingerprinting dataset. We show that for all three datasets, the narrowband OEO-based RC demonstrates competing accuracy with much less training data comparing to CNNs, and achieves an accuracy as high as 97%.