GAN-Based Siamese Neuron Network for Modulation Classification Against White-Box Adversarial Attacks

With the extensive application of modulation classification in the field of wireless communication, various methods of modulation classification have been proposed to improve the classification performance, where the deep learning (DL)-based modulation classification technology is gradually popular. However, recent studies have discovered that the DL-based modulation classification methods are vulnerable to white-box adversarial attacks, which can significantly degrade the classification performance by injecting imperceptible perturbations into clean signals. In this paper, to defend against white-box adversarial attack, we propose a generative adversarial network (GAN)-based Siamese neuron network (GSNN) for accomplishing modulation classification, where a generator is trained to generate perturbations that can easily fool the discriminator and the discriminator tries to classify correctly both original and adversarial samples generated by the generator. The discriminator acted by a Siamese network is designed to measure the similarity of sample pair based on Euclidean distance, transferring the classification problem into pairwise comparison problem. After the training of GSNN is completed, we adopt the distance-based NCM classifier to accomplish the classification task. We verify the defense performance of our proposed GSNN on both the public dataset and simulated dataset. Simulation results illustrate that our proposed method is superior to the common SoftMax classifier.