SEMI-SUPERVISED ADVERSARIAL AUTOENCODER BASED SPECTRUM SENSING FOR COGNITIVE RADIO

Spectrum sensing is a basic supporting technology of cognitive radio. Most of the existing detection methods are based on the energy method and other methods, which can be simply implemented, but the detection performances are poor in the case of low signal-to-noise ratio (SNR). In order to resolve this issue, many researchers have introduced deep learning into spectrum sensing and obtained good results, but most of the algorithms are supervised learning, which require a lot of labeled training samples. In this paper, we propose a semi-supervised spectrum sensing algorithm, which is based on the semi-supervised adversarial autoencoder (SSAAE) model. Compared with the supervised algorithms, our algorithm requires only a small amount of labeled samples. Moreover, it does not require any prior information and can directly use the received signals as the network input without any preprocessing. The simulation experiment results show that the detection probability of our proposed algorithm can reach 0.9216 when SNR and the false alarm probability are -14dB and 0.1 respectively, which is much higher than classic methods and some deep learning-based methods.