Adversarial Domain Generalization Defense for Automatic Modulation Classification

Abstract

Automatic modulation classification (AMC) technology plays a vital role in the sixth generation mobile system (6G). However, deep learning (DL) based AMC models possess unexpected vulnerability against adversarial examples, which seriously affects their applications in 6G. To this end, we propose an Adversarial Domain Generalization Defense (ADGD) algorithm to improve the adversarial robustness of DL-based AMC models. Firstly, we sequentially pre-train two classifiers that classify the original signals and the adversarial examples respectively. Secondly, we extract task-relevant features of the original signals and adversarial examples and align them. Finally, we use adversarial training to enhance the adversarial robustness of the models. The comparative experiments with various defense algorithms under white-box and black-box conditions of various attack algorithms demonstrate the outstanding defense performance of the ADGD algorithm. The proposed solution is of great significance to promote the application of AMC technology in 6G.