Robust Wideband Spectrum Sensing for Class Balance

In order to solve the spectrum scarcity problem, this paper proposes a wideband spectrum sensing method, which considers the wideband spectrum sensing as a multi-label classification task, aiming to enable wireless users to identify the idle spectrum efficiently. Firstly, the attention module is used to obtain the effective feature channel, and the effective features of the input signal are abstractly extracted using a convolutional neural network, and then the effective features are processed with temporal information by a long and short-term memory network, and finally the multi-label classification prediction of the input data is obtained. Meanwhile, asymmetric loss is introduced in the training to offset the class unbalance between occupied and idle frequency bands, which significantly enhances the robustness of the model against gradient-based attacks. Simulation experiments show that the method in this paper achieves more than 95% micro-average recall in low SNR scenarios and maintains a performance degradation of less than 3% under adversarial conditions.