CNR-SenseNet: Communication-Native Deep Learning for Robust Spectrum Sensing

Abstract

Spectrum sensing is a fundamental function in cognitive radio networks (CRNs), enabling secondary users to opportunistically access underutilized spectrum without causing harmful interference to primary users. Conventional sensing methods, such as energy detection, rely on handcrafted statistical features and often degrade under low signal-to-noise ratio (SNR) and noise uncertainty. This paper proposes CNR-SenseNet, a communication-native deep learning framework for robust spectrum sensing. Instead of treating wireless signals as generic sequences, the proposed method embeds communicationdomain priors through a lightweight multi-branch architecture that jointly exploits raw waveform information, local energy statistics, and autocorrelation structural features. To improve generalization, an augmented dataset is constructed by combining the RadioML benchmark with synthetically generated signal and noise samples under diverse SNR conditions. Experimental results show that CNR-SenseNet improves low-SNR robustness and achieves a favorable trade-off among detection performance, false alarm control, and model complexity, making it suitable for resource-constrained communication systems.