Noise-Tolerant, Deep-Learning-Based Radio Identification with Logarithmic Power Spectrum

Abstract

Wireless devices have trivial individual differences due to hardware imperfections. Therefore, extracting unique features from the physical waveforms of wireless signals enables us to identify transmitter devices. It has often been assumed that high signal-to-noise ratio (SNR) waveforms can be obtained for identification in previous studies. However, there have been a small number of evaluation examples considering lower SNR environments. Classification accuracy could decrease when the feature is extracted from signals in a low SNR. In this paper, we propose a radio identification method using the logarithmic power spectral density (PSD) with a convolutional neural network (CNN) to improve identification performance in a low SNR environment. The simulation results based on indoor experiments reveals that the proposed method achieves 90 percent accuracy even at a low SNR condition of 7 dB, which is 11 dB lower than the conventional one. Additionally, the outdoor experiment results show that the proposed method achieves 99.4 percent accuracy, which is 6.5 points higher than the conventional one.