基于特征融合的电磁信号对抗样本检测方法  

An electromagnetic signal adversarial sample detection method based on feature fusion

作　　者：徐东伟 蒋斌[2] 陈嘉峻 宣琦 王巍 赵文红[4] 杨小牛[2,3] XU Dongwei;JIANG Bin;CHEN Jiajun;XUAN Qi;WANG Wei;ZHAO Wenhong;YANG Xiaoniu(Research Center of Electromagnetic Space Security,Binjiang Institute of Artificial Intelligence,Hangzhou 310051,China;Institute of Network Security,Zhejiang University of Technology,Hangzhou 310014,China;The 36th Research Institute of China Electronics Technology Group Corporation,Jiaxing 314000,China;Jiaxing University,Jiaxing 314000,China)

机构地区：[1]杭州市滨江区浙工大人工智能创新研究院,杭州310051 [2]浙江工业大学网络安全研究院,杭州310014 [3]中国电子科技集团公司第三十六研究所,嘉兴314000 [4]嘉兴学院,嘉兴314000

出　　处：《电波科学学报》2024年第5期926-933,共8页Chinese Journal of Radio Science

基　　金：国家自然科学基金联合重点基金(U21B2001);嘉兴南湖学院校级科研项目(62211ZL)。

摘　　要：针对电磁信号调制识别智能模型容易受到对抗样本攻击的问题,提出了一种基于特征融合的电磁信号对抗样本检测方法。该方法首先通过变分模态分解对测试样本进行去噪得到去噪后的电磁信号样本,然后分别将去噪前后的电磁信号样本输入到神经网络模型中,接着计算去噪前后模型输出向量的余弦相似性值和置信度差值,最后将两个特征进行融合,输入到一个神经网络模型中进行检测。与基线方法相比,该方法在实验中取得了更高的检测成功率。本文方法具有时间复杂度低、易于实现的优点,为电磁信号调制识别智能模型提供了一种新颖的对抗样本检测方法。Aiming at the problem that the intelligent model of electromagnetic signal modulation recognition is vulnerable to adversarial samples,a feature fusion electromagnetic signal adversarial sample detection method is considered.At first,the electromagnetic signal samples after denoising are obtained through variational mode decomposition,and then the electromagnetic signal samples before and after denoising are sent into the neural network model,and then the cosine similarity value and confidence difference value of the model output vector before and after denoising are calculated.Finally,the two features are fused and sent into a neural network model for detection.Compared with the baseline method,this method has higher detection success rate in the experiment.This method has the advantages of low time complexity and easy implementation,and provides a novel adversarial sample detection method for the intelligent model of electromagnetic signal modulation recognition.

关 键 词：电磁信号调制识别 对抗样本检测 特征融合 余弦相似性 置信度差 

分 类 号：TN975[电子电信—信号与信息处理]