Prototype-Guided Anti-Jamming Framework Against White-Box Adversarial Attack in Radio-Frequency Fingerprint Identification

Radio-frequency fingerprint (RFF) is a promising solution for physical layer security authentication. Existing schemes of radio-frequency fingerprint identification (RFFI) leverage unique hardware impairments as device identifiers, which can be separated from noisy signals using deep learning (DL) methods. However, the non-interpretable characteristics and vulnerabilities of neural networks render current DL-based RFFI methods susceptible to adversarial attacks. Conventional adversarial training (AT) methods against the attacks also face challenges of low convergence speed and significant performance degradation on unperturbed samples. To address these limitations, we design a prototype-guided anti-jamming framework for robust RFF extraction under white-box untargeted adversarial attacks. Specifically, a denoising network using multi-level wavelet convolutional neural network (MWCNN) is proposed to speed up convergence in low signal-to-noise ratio (SNR) scenarios on adversarial samples. For feature alignment in hidden layers of the classification model, a prototype-guided training method is used in our framework. The defense performance of our framework is evaluated using a real-world dataset of WiFi preamble signals. Our results demonstrate that the proposed framework outperforms the conventional AT method in classification accuracy and stability even under strong adversarial perturbations.