SENSITIVITY ANALYSIS OF A DEEP LEARNING-BASED SPECTRUM SENSING MODEL FOR COGNITIVE RADIO NETWORKS
ABSTRACT The deployment of cognitive radio (CR) networks for next-generation wireless systems is critically hindered by the stringent requirements for reliable spectrum sensing in low Signal-to-Noise Ratio (SNR) regimes, particularly for Edge devices with constrained computational resources. While deep learning (DL) has emerged as a powerful paradigm for spectrum sensing, the complex interplay between architectural design, hyperparameter tuning, and operational efficiency under noisy conditions remains largely unexplored. This paper presents a comprehensive systematic evaluation of 16 hybrid deep learning models, spanning two distinct architectural families (Model95 and Model128) with eight variants each, to ascertain their robustness and reliability for spectrum sensing at a challenging -5 dB SNR. Through a rigorous sensitivity analysis, we systematically investigate the impact of critical network parameters, including depth, learning rate, batch size, and input dimensionality, on detection performance. Our empirical findings reveal a non-linear dependence of performance on these parameters, underscoring the necessity of tailored optimization rather than generic architectural scaling. Among all contenders, Model95 Variant 2 emerges as the definitive optimal Edge-deployable solution, achieving a superior probability of detection of 90%, an F1-score of 81.80%, and a classification accuracy of 75%, while maintaining an exceptionally lean parameter footprint of merely 1.1 million. This unique combination of high perceptual performance and excellent computational efficiency establishes Model95 Variant 2 as a paradigmatic benchmark for practical CR implementations. The results demonstrate that strategic architectural pruning and hyperparameter fine-tuning can yield near-optimal sensing capabilities at the edge, effectively bridging the gap between theoretical DL potential and real-world, low-latency cognitive radio applications.
Authors : Pallam, S.W., Thuku, I.T., Luka, M.K., Visa, M.I., Abana, M.A., Uyoata, E.U. and Zummo, A.H.
Category : Open Access Volume (Issue) : 12(1) Date Uploaded : 5th August 2026