Deep Neural Network Architectures for Spectrum Sensing Using Signal Processing Features

In this work, we consider a performance comparison of deep learning-based approaches to the problem of spectrum sensing (SS) in cognitive radios. Towards this end, we use signal processing (SP) features such as energy, differential entropy, geometric power and p-norm. For the classification problem of SS, we employ deep neural network (NN) architectures such as multi-layer perceptron (MLP), convolutional NN, fully convolutional network, residual NN (ResNet), long short-term memory and temporal convolutional network. Through extensive experiments based on real-world captured datasets and Monte Carlo simulations, we show that MLP and ResNet architectures offer the best performance in terms of probability of detection, for a given predefined level of probability of false-alarm. Further, we show that NN architectures trained with a combined set of the SP features yield the best performance.