Robust Automatic Modulation Classification Using Domain-Adversarial Neural Network with Data Inconsistency

Automatic modulation classification (AMC) is an important wireless communications technology. With the rapid development of deep learning (DL), DL based AMC has been widely used because of its powerful classification ability. Previous research on DL based AMC usually assumes that the data distribution in the training and inference phases is consistent. However, in practical applications, the uncertainty of the communications environment often leads to inconsistent data distributions between training and inference, and resulting in the decrease of classification accuracy. To combat the problem, this paper proposes a domain-adversarial neural network based modulation recognition algorithm. The proposed algorithm uses classifiers and domain classifiers for adversarial training to enable the feature extractor to extract features that have both class-specificity and domain-inconsistency. Experimental results show that the algorithm can effectively reduce the side effects of data inconsistency caused by channel variations, improving the model's generalization and robustness.