Title:Real-World Considerations for Deep Learning in Wireless Signal Identification Based on Spectral Correlation Function

Abstract:This paper proposes a convolutional neural network (CNN) model which utilizes the spectral correlation function (SCF) for wireless radio access technology identification without any prior information about bandwidth and/or the center frequency. The sensing and classification methods are applied to the baseband equivalent signals. Two different approaches are elaborated. The proposed method is implemented in two different settings; in the first setting, signals are jointly sensed and classified. Sensing and classification are conducted in a sequential manner in the second setting. The performance of both approaches is discussed in detail. The proposed method eliminates the threshold estimation processes of classical estimators. It also eliminates the need to know the distinct features of signals beforehand. Over-the-air real-world measurements are used to show the robustness and the validity of the proposed method and various wireless signals are successfully distinguished from each other without any a priori knowledge. The over-the-air real-world measurements are also shared in the format of SCF. The performance of SCF-based identification is compared with the cases when fast Fourier transform and amplitude-phase representation are used as the training inputs for CNN. The comparative performance of the proposed method is quantified by precision, recall, and F1-score metrics. Moreover, a setup to compare the performance of the proposed approach with classical cyclostationary features detection (CFD) is prepared. Measurement results indicate the superiority of the proposed method against CFD, especially at the low signal-to-noise ratio regime.