Performance of Radio Frequency Fingerprinting : Parameters and Models

Radio Frequency fingerprinting (RFF) is emerging as a viable alternative to authenticating radio devices, serving as a mitigation technique for spoofing and impersonation attacks on the wireless channel. RFF is rooted in the fact that each radio transducer features a distinctive radio fingerprint that is impractical, or even impossible, to forge with any other device. In this work, We conduct an extensive inspection of the analysis of current state-of-the-art RFF approaches by comparing deep learning techniques and the associated methodologies. We consider real measurements in a controlled scenario and compare different configurations and classifiers in terms of performance and training time. Our findings show that the performance of the 11 classifiers considered is deeply biased by the methodology considered during the selection of the data for the training and testing datasets. Training and testing on different measurements or when radios are power-cycled significantly affect the accuracy of the classifier. Overall, our investigation sheds light on the best practices and configurations to consider to maximize the performance of RFF systems deployed in the wild.