Semi-supervised Modulation Recognition Greatly Improved by Strong Data Augmentation

Modulation recognition plays an important role in modern wireless communications. Recent work has shown that modulation recognition based on deep learning significantly outperforms conventional approaches. However, this superiority relies largely on using plenty of labeled data for supervised learning, whereas training deep neural networks with limited data generally falls into overfitting, resulting in poor performance. In addition, it is also challenging to obtain plenty of labeled data in real-world communication activities, with expensive and time-consuming costs. To this end, we present a semi-supervised method for modulation recognition, which can take advantage of unlabeled samples that are more easily accessible in practice to enhance generalization and thus reduce such demand for labeled data. By introducing strong data augmentation, we improve supervised training and simultaneously perturb unlabeled data for consistency-based regularization, resulting in a remarkable generalization improvement. Experimental results on RadioML 2018.01A dataset demonstrate that our proposed method for semi-supervised modulation recognition is far superior to other competing ones and achieves almost fully supervised performance with a very small number of labeled samples.