调制识别中目标对抗攻击

由于深度学习算法具有特征表达能力强、特征自动提取以及端到端学习等突出优势，因此被越来越多的研究者应用至通信信号识别领域。然而，对抗样本的发现使得深度学习模型极大程度地暴露在潜在的风险因素中，并对当前的调制识别任务造成严重的影响。本文从攻击者的角度出发，通过对当前传输的通信信号添加对抗样本，以验证和评估目标对抗样本对调制识别模型的攻击性能。实验表明，当前的目标攻击可以有效地降低模型识别的精确度，所提出的logit指标可以更细粒度地用于衡量攻击的目标性效果。

Targeted adversarial attack in modulation recognition

Since deep learning algorithms have outstanding advantages such as strong feature expression ability, automatic feature extraction, and end-to-end learning, more and more researchers have applied them to the field of communication signal recognition. However, the discovery of adversarial examples exposes deep learning models to potential risk factors to a great extent, which has a serious impact on current modulation recognition tasks. From the perspective of an attacker, adversarial examples are added to the currently transmitted communication signal to verify and evaluate the attack performance of the target countermeasure sample to the modulation recognition model. Experimental results show that the current targeted attack can effectively reduce the accuracy of model recognition, and the constructed logit indicator can be better applied to measure the targeted effect more fine-grained.