Multimodel Selection and Computation Resource Allocation Driven Cooperative Spectrum Sensing

Cooperative spectrum sensing (CSS) plays a crucial role in this era of explosive Internet of Things with scarce spectrum resources, since it can effectively enhance the sensing accuracy with the cooperation of secondary users (SUs). However, most existing CSS algorithms primarily focus on increasing the cooperative detection accuracy, while neglecting the computational complexity or sensing latency. Therefore, we propose a deep learning (DL) driven CSS scheme with the consideration of dynamic multimodel selection and suitable resource allocation. Specifically, we first derive the closed-form expressions to fit and characterize the detection and false alarm probabilities of three popular DL models, including the convolutional neural network, the long short-term memory (LSTM) network and the hybrid convolutional LSTM network. Then, the problem of minimizing the cooperative sensing error is formulated under the constrains of limited computational resource and sensing latency. Finally, the cross-entropy algorithm is employed to dynamically select the most suitable cooperating SU set and their correspondingly matched models, to balance the sensing accuracy and computational complexity. Simulation results demonstrate that our CSS scheme are much more robust and computationally efficient compared to some well-known CSS algorithms, especially in achieving extremely high sensing accuracy at low transmit power or low received signal-to-noise ratio.