Triplet Network and Unsupervised-Clustering-Based Zero-Shot Radio Frequency Fingerprint Identification With Extremely Small Sample Size

By exploiting the inherent hardware characteristics of wireless devices, radio frequency fingerprint identification (RFFI) has been widely applied in device authentication and spoofing attack detection to improve the security. However, due to the dependence on the training sample size, the state-of-the-art deep learning (DL)-based identification methods will face serious overfitting problem with inadequate training samples. Besides, in noncooperative scenarios, most existing methods cannot tackle the challenge of zero-shot identification, i.e., identifying the objects outside of the training data with no prior samples, which obstructs their practical applications. To solve these two problems, in this article, we propose a novel identification method that combines the deep neural network (DNN) and the unsupervised clustering. In this design, after offline training, the triplet loss convolutional neural network (TLCNN) can be utilized to extract the features of the radio frequency signals outside of the training set. Then, the
$K$
-means++ clustering algorithm is applied to the extracted features to realize zero-shot RFFI. Moreover, to improve the identification performances of the proposed design under small sample conditions, the random integration (RI) augmentation and the variational mode decomposition (VMD) are exploited to preprocess the input signals, and a
$K$
-value estimation method for the
$K$
-means++ clustering is proposed to ensure the effectiveness of the proposed design. Experiment on digital mobile radio (DMR) portable radios validates that the proposed design can achieve the best identification performances and the minimal time and space costs compared with the benchmark schemes.