Adversarial Attacks on Deep Neural Networks Based Modulation Recognition

Abstract

Modulation recognition models based on deep neural network (DNN) have the advantages of automatic feature extraction, fast recognition and high accuracy. However, due to the interpretability defects, DNN models are vulnerable to adversarial examples designed by attackers. Most existing researches focus on the accuracy of modulation recognition models, while ignoring the huge threat of adversarial examples to the safety and reliability of the models. In the field of modulation recognition, many existing attack methods have good attack performance for simple neural networks, but poor performance for more complicated DNNs. Therefore, this paper proposes an adversarial attack method based on dynamic iterative. The proposed method uses a dynamic iterative step that changes with iteration instead of being fixed. Simulation results show that the proposed attack method has better attack performance when the disturbance is specified than the traditional attack methods.