Deep learning (DL) models are vulnerable to adversarial attacks, by adding a subtle perturbation which is imperceptible to the human eye, a convolutional neural network (CNN) can lead to erroneous results, which greatly reduces the reliability and security of the DL tasks. Considering the wide application of modulation recognition in the communication field and the rapid development of DL, by adding a well-designed adversarial perturbation to the input signal, this article explores the performance of attack methods on modulation recognition, measures the effectiveness of adversarial attacks on signals, and provides the empirical evaluation of the reliabilities of CNNs. The results indicate that the accuracy of the target model reduce significantly by adversarial attacks, when the perturbation factor is 0.001, the accuracy of the model could drop by about 50 ${\%}$ on average. Among them, iterative methods show greater attack performances than that of one-step method. In addition, the consistency of the waveform before and after the perturbation is examined, to consider whether the added adversarial examples are small enough (i.e., hard to distinguish by human eyes). This article also aims at inspiring researchers to further promote the CNNs reliabilities against adversarial attacks.

中文翻译：

---

卷积神经网络在调制识别中的对抗性攻击

深度学习（DL）模型容易受到对抗攻击，通过添加人眼无法察觉的微妙扰动，卷积神经网络（CNN）会导致错误结果，从而大大降低DL任务的可靠性和安全性。考虑到调制识别在通信领域的广泛应用和DL的快速发展，本文通过对输入信号添加精心设计的对抗性扰动，探讨了攻击方法在调制识别中的性能，衡量了对抗性攻击对攻击的有效性。信号，并提供对CNN可靠性的经验评估。结果表明，当干扰因子为0.001时，对抗性攻击会大大降低目标模型的精度， $ {\％} $一般。其中，迭代方法比单步方法具有更高的攻击性能。另外，检查摄动前后的波形一致性，以考虑所添加的对抗示例是否足够小（即，人眼难以辨别）。本文还旨在激励研究人员进一步提高CNN对抗对抗性攻击的可靠性。