Toward Novel Time Representations for RFF Identification Using Imperfect Data Sets

As an inherent attribute of hardware circuit, radio-frequency fingerprint (RFF) is hardly forged and unique. Recently, the connection with deep learning has made it one of the most powerful guarantees of physical-layer security. Most existing RFF-based methods are designed under ideal data sets, thus tend to be less versatile in real-world scenarios. To address this problem, we propose a novel RFF identification scheme toward imperfect data sets of small sample size and fragmentary signal. Two novel time representations are proposed to visualize the RFFs by digging the embedded temporal information, which are named 1-D Gramian angular fields (1DGAF) and 2-D Gramian angular fields (2DGAF), respectively. Specifically, extremely short segments of the received preamble are transformed into images using 1DGAF and 2DGAF. Then, the images are fed into a channel-selectable convolutional neural network (CNN) for further identification. Theoretical analysis indicates that the proposed methods can augment the separability of the original data, leading to a better identification performance. Experimental results show that the accuracy reach 94.81% with only three half-sine waves and 99.26% with a quarter of the preamble at the signal-to-noise ratio level of 30 dB. The robustness of the proposed methods using unseen symbols has also been verified.