An improved metric-active learning approach for few labeled radio frequency fingerprinting

Abstract

Radio frequency fingerprinting (RFF) is an effective non-cryptographic method for physical-layer authentication. Current deep learning (DL) approaches have achieved strong results in RFF identification but typically require large annotated datasets, making data collection and labeling costly. To address this, we propose a novel Active Learning (AL) framework that builds a robust classifier with minimal labeled data through incremental learning. Our framework improves AL in two ways: (1) using a Siamese-based metric learning model to capture discriminative features from data-pair similarities, and (2) adopting a cost-effective sample selection strategy that reduces manual labeling while enhancing accuracy. Unlike methods that focus only on low-confidence samples, our approach also leverages high-confidence samples from the unlabeled pool, assigning them pseudo-labels to expand the training set. Experiments on Laboratory LoRa devices show that the framework achieves superior performance with fewer labeled samples.