通信信号智能调制识别对抗攻击研究  

Research on intelligent modulation and recognition of communication signals to counter attacks

作　　者：张正豪 陈家军 黄知涛[1] 王翔 柯达 Zhang Zhenghao;Chen Jiajun;Huang Zhitao;Wang Xiang;Ke Da(State Key Laboratory of Compex Eectromagnetic Environment Effects on Eectronics&Information,System,National University of Defense Technology,Changsha 410073,Hunan,China;School of Elec-tronic Countermeasures,National University of Defense Technology,Hefei 230000,Anhui,China;Unit 73676 of PLA,Jiangyin 214400,Jiangsu,China)

机构地区：[1]国防科技大学电子科学学院电子信息系统复杂电磁环境效应国家重点实验室,湖南长沙410073 [2]国防科技大学电子对抗学院,安徽合肥230000 [3]中国人民解放军73676部队,江苏江阴214400

出　　处：《航天电子对抗》2023年第5期11-16,共6页Aerospace Electronic Warfare

摘　　要：基于深度学习技术的通信信号调制识别算法由于其优秀的特征提取能力,极大地提升了调制识别任务的精度,但是对抗样本特性的存在,导致基于深度学习的调制识别模型的安全性受到了极大威胁,通过在训练好的调制识别网络中添加设计好的特定微小扰动,就可以使得调制识别模型完全失效。研究了基于深度学习的调制识别模型及其对抗样本攻击方法,提出一种基于快速梯度符号法的定向扰动生成方法,该方法在扰动和原始信号功率比为-21 dB的条件下,针对11类常见的不同调制种类的通信信号生成扰动,实现对通信信号调制识别模型的定向攻击,为智能调制识别模型的攻防对抗提供参考。Communication signal modulation recognition algorithms based on deep learning technology have greatly improved the accuracy of modulation recognition tasks due to their excellent feature extraction ability.However,due to the presence of adversarial sample characteristics,the security of modulation recognition models based on deep learning has been greatly threatened.By adding designed specific small perturbations to the trained modulation recognition network,the modulation recognition model can be completely invalidated.A modulation recognition model based on deep learning and its methods for countering sample attacks are studied.A directed disturbance generation method based on fast gradient symbol method is proposed.This method generates distur-bances for 11 common types of communication signals with different modulation types under the condition of a power ratio of-21 dB between the disturbance and the original signal,achieving directed attacks on the modula-tion recognition model of communication signals.It provides reference for the attack and defense ends of the intel-ligent modulation recognition model.

关 键 词：深度学习 调制识别 对抗样本 

分 类 号：TN975[电子电信—信号与信息处理]