Modulation Recognition Based on Deep Co-Training

Abstract

While deep learning has significantly improved signal modulation recognition performance, these algorithms need a large number of labeled samples for training. But in real-world communication conditions, a large number of unlabeled signal samples is often more easily accessible. To address this problem, we propose a semi-supervised approach based on Deep Co-Training that maximizes the utilization of unlabeled data. We first augment the signal samples and initialize two different CLDNN network by pre-training. Then, we construct multiple views using the gradient attack algorithm and measure the consistency of the outputs with Jensen-Shannon Divergence. The simulation findings indicate that the strategy outperforms supervised learning under limited sample conditions, improving recognition accuracy by 5.75% to 11.01%.