A Dual Transfer Framework for Cooperative Spectrum Sensing in Cognitive Radio

Introducing machine learning (ML) into cooperative spectrum sensing (CSS) in cognitive radio has yielded effective data-driven solutions for the spectrum shortage problem. Significant quantities of labeled data are required for ML-based CSS model training. However, collecting and annotating data repeatedly under dynamic spectrum environmental conditions is time-consuming, costly, and impractical. To this end, we propose a novel non-parametric dual transfer framework for CSS (DTCSS) to solve the poor generalization sensing performance caused by insufficient labeled data in the target environment with different wireless signals and propagation. DTCSS features a unique design that employs a two-stage learning approach. The offline training stage aims to transfer domain-level and class-level knowledge from the existing environment to the target environment and train a target detector. The objective of the online sensing phase is to utilize the trained detector to deduce the spectrum status of the target environment. DTCSS is robust and effective without hyperparameter tuning. Results from simulations indicate that DTCSS can achieve competitive sensing performance.