STC-RFF: Self-Supervised Radio Frequency Fingerprinting via Spatiotemporal Consistency

Radio frequency (RF) fingerprints originate from the unique hardware characteristics of devices during signal transmission and serve as effective, nonreplicable identifiers for device authentication. While deep learning demonstrates significant potential in RF fingerprinting (RFF) and both semi-supervised and self-supervised methods can address the challenge of large-scale RF signal annotation by leveraging unlabeled data, existing approaches primarily focus on deep learning methodology while overlooking the inherent characteristics of RF fingerprints. In this article, we propose a self-supervised RFF method based on the spatiotemporal consistency of RF fingerprints (STC-RFFs). First, we discuss the spatiotemporal consistency (STC) of RF fingerprints and develop a data augmentation based on STC to enhance view generation while preserving semantic consistency across different views. Subsequently, an efficiency contrastive loss function is employed to optimize the feature encoder by utilizing multiple pooling to facilitate hierarchy representation learning. Our method achieves the identification accuracy of 84.86%, 89.85%, and 92.70% with 5%, 10%, and 20% labeled data, respectively, when evaluated on a real-world ADS-B dataset. The results outperform recently advanced self-supervised methods at 5% (71.70%), 10% (85.70%), and 20% (91.20%) labeled ratios and surpass semi-supervised techniques at 5% (79.70%). Meanwhile, the proposed method demonstrates robust performance across various labeled ratios.