Investigating Sparse Neural Networks for Radio Frequency Fingerprint Identification

Radio-Frequency Fingerprint Identification (RFFI) shows promise for enhancing wireless identification security through unique emitter imperfections. However, this method, which primarily relies on deep learning, faces challenges for a real-time deployment on edge devices. Both memory and computational resources are indeed presented as the major obstacle of such deployment. To address these issues, this paper proposes to investigate the behavior of compressed convolutional neural networks when using unstructured pruning in a data-free scenario, on several public datasets. Indeed, unstructured pruning exhibits significant compression capabilities while preserving performance with minimal computation. We show that state-of-the-art RFFI neural networks can be pruned by up to 70% of the weights, while maintaining F1-Scores above 99%, without retraining. Additionally, we advocate the use of data from a later day, referred to as resilience, as an additional indicator of network performance.