A transfer learning-based GAN for data augmentation in automatic modulation recognition

Abstract The automatic modulation recognition (AMR) algorithms based on deep learning (DL) have achieved high classification accuracy by automatically extracting deep features from massive data. However, in real-world scenarios, sufficient training data is always difficult to collect, which affects the performance of DL-based models. As a type of data augmentation algorithm, Generative Adversarial Networks (GANs) can generate artificial data similar to the given real data and thus solve the problem of insufficient data, whereas the training process of GANs is also affected by limited number of data samples. Inspired by the successful application of transferring GANs in the field of image generation, this paper employs transfer learning-based GANs to enlarge the training data by generating the constellation diagram images of radio signals, which can effectively solve the problems of divergence and model collapse. We feed the enhanced dataset into a CNN model for modulation recognition and the experimental results demonstrate that our proposed method achieves a performance improvement ranging from 1% to 5.1% compared with the result of the original limited training data.