Radio Classify Generative Adversarial Networks: A Semi-supervised Method for Modulation Recognition

We introduce Generative Adversarial Network (GAN) into the radio machine learning domain for the task of modulation recognition by proposing a general, scalable, end-to-end framework named Radio Classify Generative Adversarial Networks (RCGANs). This method naively learns its features through self-optimization during an extensive data-driven GPU-based training process. Several experiments are taken on a synthetic radio frequency dataset, simulation results show that, compared with some renowned deep learning methods and classic machine learning methods, the proposed method achieves higher or equivalent classification accuracy, superior data utilization, and presents robustness against noises.