Deep Learning of Radio Frequency Fingerprints from Limited Samples by Masked Autoencoding

Radio frequency fingerprints (RFFs) refer to the unique characteristics of signals transmitted by each emitter, which are valuable for physical layer security. Despite the fact that deep learning based methods have shown significant advantages for RFF extraction, they generally require large amount of data to avoid overfitting. To improve the performance of deep learning based RFF extraction methods with limited training samples, we propose an unsupervised pre-training method based on masked autoencoding (MAE) for RFF learning. Specifically, the neural network for RFF learning is first pre-trained to predict the masked segment of signals, and then fine tuned supervisedly with identity labels. Effects of different masking patterns for RFF learning are also evaluated. Experimental results of both simulated dataset and real dataset show that MAE pre-training with block-wise channel-aligned masking performs better than vanilla training method under various conditions, requiring no additional training data.