CCD-GAN for Domain Adaptation in Time-Frequency Localization-Based Wideband Spectrum Sensing

In practical spectrum sensing scenarios, sample distribution in the training dataset, i.e., source spectrum domain, is generally different from that of the test dataset, i.e., target spectrum domain, which results in the domain adaptation problem. For the emerging time-frequency localization (TFL) based wideband spectrum sensing, we propose a Consistency Constrained Dual Generative Adversarial Network (CCD-GAN) to address this problem. To achieve multi-level feature alignment in the TFL task, we introduce consistency constraints into a dual GAN, which takes into account domain consistency, content consistency, and TFL distribution consistency. Moreover, we retrain a lightweight detector and use transfer learning to improve the retraining efficiency. Finally, experimental simulation proves the effectiveness of CCD-GAN in solving domain adaptation problems, and shows the improvement of generalization ability and convergence speed brought by transfer learning.

Automated summary

In this paper, we propose a Consistency Constrained Dual Generative Adversarial Network (CCD-GAN) to address the domain adaptation problem in practical spectrum sensing scenarios. By introducing consistency constraints and combining with retraining a lightweight detector and transfer learning, the proposed method achieves multi-level feature alignment in the time-frequency localization (TFL) task and shows improvement in generalization ability and convergence speed.