Spectrum sensing based on deep convolutional generative adversarial networks

As the basis of cognitive radio technology, Spectrum sensing (SS) has received widespread attention because it is very important to improve spectrum efficiency. However, the limited sensing time makes it difficult to obtain sufficient sample data, which will seriously affect the performance of the spectrum sensing model. In this paper, SS is considered to be a binary classification problem, in which Deep Convolutional Generative Adversarial Networks(DCGAN) is improved and used to expand the training set to cope with the shortage of sample data. More specifically, the sampling covariance matrix of the received signal is firstly transformed into the true color picture which is divided into a training set and a test set. After that, the obtained training set is expanded with the improved DCGAN. Finally, the LeNet network is trained based on the extended data. Simulation results show that the proposed scheme greatly improve the sensing accuracy. The probability of detection(PD) and the probability of false alarm(PFA) fluctuate less after expanding the dataset with DCGAN. Especially when SNR= −4dB, the minimum value of PD increases by 0.2.