Advancing Cognitive Radio Spectrum Sensing with Deep Learning and Explainable AI: Cross-Domain Attribution Reveals SNR-Dependent Feature Bias

Abstract

The growing demand for wireless spectrum has made Dynamic Spectrum Access through Cognitive Radio a necessity, its core function is spectrum sensing, remains unre-liable when traditional methods like Energy Detection face low Signal-to-Noise Ratio (SNR) environments. The proposed work presents a deep learning framework for binary spectrum sensing that addresses three open challenges: robust detection at low SNR, computational efficiency for edge deployment, and decision transparency through Explainable AI. Three architectures are implemented and compared under identical conditions on the RadioML dataset (320,000 frames, 8 digital modulations, -20dB to +18 dB SNR): a Convolutional Neural Network operating on Continuous Wavelet Transform scalograms (1.2M parameters), a Bidirectional GRU with attention operating on raw I/Q sequences (101K parameters), and a Hybrid CNN-LSTM (78K parameters). All models are trained on unit-energy-normalized data where the classical Energy Detector is reduced to random chance (Pd = 0.03). The Bi-GRU achieves the highest accuracy of 87.79% with 3.5 ms inference latency outperforming the CNN by 2.44 percentage points (p; 0.001) while using 12× fewer parameters. A cross-domain Explainable AI analysis applying 2D Grad-CAM and Integrated Gradients reveals that spatial attribution on CWT representations is strongly SNR-dependent, while temporal attribution on I/Q remains fixed regardless of signal strength. A circular-shift control experiment confirms that observed attention patterns reflect dataset positional artifacts rather than genuine signal features. The results establish that temporal processing of raw I/Q sequences provides superior detection performance with greater efficiency than spatial time-frequency analysis for short-frame spectrum sensing.