Performance Analysis of a Deep Neural Network-Based Spectrum Sensing Approach

Radio spectrum is a limited resource and the traditional fixed spectrum allocation methods which permit only the Primary Users (PUs) to access the allocated spectrum, even when it is idle, will quickly deplete the radio resources. The growing network demands and emerging 6G applications, like Intelligent Healthcare, call for the effective utilization of the underutilized spectrum. Dynamic Spectrum Management (DSM) is a spectrum allocation approach which with the help of Spectrum Sensing (SS) detects and transmits in the underutilized frequency bands. Deep Neural Networks (DNNs)-based SS techniques being data-driven can enhance the performance of 6G communication systems by eliminating the drawbacks of conventional SS approaches which suffer from missed detection of the PU, and traditional Machine Learning (ML) techniques which need manual feature extraction. This article explores the publicly available RadioML 2016.10A dataset and demonstrates the use of a Deep Learning (DL)-based SS approach on a dataset formed by using the RadioML 2016.10A samples and synthetically generated noise. SS is considered as a binary classification problem and the proposed deep autoencoder-based technique reports a 97.41% classification accuracy on the entire dataset. The performance of the model is analyzed for various Signal-to-Noise Ratio (SNR) values and the model achieves high classification accuracies. The proposed algorithm when analyzed for different modulation schemes reports a 100% test accuracy for 8-Phase-Shift Keying (8PSK), Binary Phase-Shift Keying (BPSK), Quadrature Amplitude Modulation-16 (QAM16), and Pulse Amplitude Modulation-4 (PAM4) schemes.