Explainable RFF: Radio Frequency Fingerprint via Spectrogram Analysis

Abstract

Radio Frequency Fingerprint (RFF) is a unique characteristic of radio signals impacted by hardware imperfections of the device’s radio front-end. This characteristic can be used as a security measure to identify individual devices. There are various methods to extract RFF features, ranging from statistical analysis to deep learning methods, which have seen increasing adoption among researchers, given their automatic feature extraction capabilities and performance. However, the black-box nature of deep learning models is a significant drawback, as it raises concerns about the model’s reliability, especially in critical applications such as security. In this paper, we use In-phase and Quadrature (I/Q) samples to create spectrograms and train an EfficientNet model for device classification. The model is trained on a Wi-Fi dataset of 20 devices, achieving 99% accuracy on the test set. More importantly, to understand the model’s decision-making process, we apply the Gradient-weighted Class Activation Map (Grad-CAM) to visualize its attention. Grad-CAM highlights specific spectrogram regions that are linked to the device’s fingerprint signature, including power leakage, Direct Current (DC) components, and energy patterns.