Explainable Efficiency: Grad-CAM Analysis of Image-Based Radio Frequency Fingerprinting

Radio Frequency Fingerprinting (RFF) is a physical (PHY) layer security technique enabling the identification of a radio transmitter without resorting to shared secrets while combining deep learning and signal processing techniques. State-of-the-art scientific contributions focus on improving identification performance in different scenarios and configurations while considering the RFF methodology as a black-box. In this work, we apply eXplainable Artificial Intelligence (XAI) techniques to expose the discriminative features allowing a neural network model to identify radio transmitters at the PHY layer. Our analysis considers image-based RFF classifiers, where received signals (in the form of IQ symbols) are processed into images, and proves that symbols contribute unequally to the identification of the transmitter. In particular, our results demonstrate that image pre-processing can be leveraged to significantly reduce the overhead of training (up to 50%) and testing (up to 70%) without affecting the classifier's final accuracy.