Robust Spectrum Sensing Using Adaptive Riemannian Attention Networks in Cognitive Radio Networks

Riemannian geometry-based methods, such as Riemannian k-nearest neighbors (R-kNN), Riemannian support vector machines (R-SVM), tangent-space models, and Riemannian convolutional neural networks (R-CNNs), have demonstrated strong effectiveness for spectrum sensing applications. These approaches exploit the geometric structure of covariance matrices to improve detection performance in cognitive radio systems. However, their performance degrades in low signal-to-noise ratio (SNR) conditions and rapidly time-varying environments. This degradation is mainly due to reliance on static Riemannian metrics, limited capability to discriminate among spectral subbands, and the high computational cost associated with repeated geodesic distance calculations. To address these limitations, this paper introduces the Adaptive Riemannian Attention Network (ARAN), an artificial intelligence–based framework for robust spectrum sensing. ARAN integrates Riemannian manifold learning on the symmetric positive definite (SPD) cone with a neural attention mechanism and adaptive metric learning in the tangent space. Instead of relying on fixed geometric assumptions, ARAN employs data-driven learning and attention-based feature reweighting to dynamically adapt the underlying Riemannian geometry as spectrum conditions evolve. This adaptive behavior improves the separability between hypotheses H₀ and H₁, even in the presence of severe noise and interference. Four controlled experiments are conducted to evaluate ARAN, including primary user detection, standalone performance analysis, comparative benchmarking against existing Riemannian models, and overall detection accuracy assessment. Results show that ARAN achieves a 5–10% improvement in detection probability at SNR values as low as −10 dB, reduces false alarm rates, and yields more favorable receiver operating characteristic curves, while maintaining manageable computational complexity.