Adversarial Learning in Transformer Based Neural Network in Radio Signal Classification

Deep Learning has attracted significant interests in wireless communication design problems. However, recent studies discovered that the deep neural network is vulnerable to adversarial attacks in the sense that a carefully designed and imperceptible perturbation to the input of the neural network could mislead the prediction of the neural network. In this paper, motivated by attractive classification performance of the transformer based neural networks, we analyse the vulnerability and robustness of the transformer against adversarial attacks in modulation classification scenarios. Using real datasets, we demonstrate that the transformer can achieve higher accuracy as compared to a convolutional neural network in the presence of adversarial attacks.