Multi-Antenna Spectrum Sensing via Deep Diffusion-Attention Networks

Abstract

Insufficient labeled training data for spectrum sensing can lead to significant performance degradation and overfitting of existing deep learning methods under low signal-to-noise ratio (SNR) conditions. This paper proposes a diffusion-based data augmentation method for multi-antenna spectrum sensing, which leverages denoising diffusion probabilistic model to generate high-quality synthetic samples. These synthetic samples combined with the limited labeled dataset are used to train a novel deep convolutional network. This network integrates depthwise separable convolutions and convolutional block attention module. A confidence-weighted pseudo-labeling approach is adopted to alleviate the negative impact of low-quality synthetic samples. Experimental results demonstrate the superiority of the proposed method, showing significant improvements in detection accuracy and robustness, especially under low SNR conditions.