Spectrum sensing enables Secondary Users (SUs) in cognitive radio networks to opportunistically access underutilized spectrum without causing interference to Primary Users (PUs). While deep neural networks have shown strong potential for this task, most existing approaches rely heavily on simulated datasets, limiting their applicability in real-world environments. In this work, we present a real-world dataset collected using two Universal Software Radio Peripheral (USRP) B210 devices, where one emulates the PU and the other acts as the SU. Additionally, a baseline additive white Gaussian noise dataset is collected using a wired setup. We further propose a lightweight convolutional neural network with dilated convolutions that achieves the highest sensing accuracy across all modulation schemes. At a signal-to-noise ratio of -6 dB, it outperforms the second-best method by 19.35% while being 40.5% smaller in size and requiring 85.18% fewer floating-point operations. Experiments using our real-world datasets and synthetic datasets derived from the widely used RadioML2016.10A and RML22 datasets reveal a significant drop in model performance on real-world data, highlighting the limitations of simulation-based evaluations and the importance of realistic datasets in spectrum sensing research.
Syed et al. (Thu,) studied this question.