基于显著图的电磁信号对抗样本生成方法  

An Electromagnetic Signal Adversarial Examples Generation Method Based on Saliency Map

作　　者：周侠 张剑[1] 李宁安 ZHOU Xia;ZHANG Jian;LI Ning-an(Wuhan Digital Engineering Institute,Wuhan,Hubei 430205,China)

机构地区：[1]武汉数字工程研究所,湖北武汉430205

出　　处：《电子学报》2023年第7期1917-1928,共12页Acta Electronica Sinica

基　　金：国防科技技术领域基金项目(A类)(No.A24011)。

摘　　要：基于深度学习的电磁信号识别模型具有高效、准确和人工干预少的优点,然而其与传统神经网络模型一样容易受到对抗样本的影响.研究对抗样本对测试和提升模型的安全性和鲁棒性有着重要意义.为生成高质量电磁信号对抗样本,本文提出了基于雅可比显著图批量特征点攻击算法(Batch Points Jacobian-based Saliency Map Attack, BP-JSMA).与传统雅可比显著图的攻击方法相比,BP-JSMA通过批量选取关键特征点能够更快生成对抗样本.此外,针对电磁信号数据的特点,增加自适应扰动限制,使得生成的对抗样本更具隐蔽性.在公开数据集的实验结果表明,与雅可比显著图攻击方法相比,BP-JSMA在生成速度方面提升了11倍,隐蔽性提升了10%;而与传统快速梯度符号攻击算法相比,攻击成功率提升了24%,隐蔽性提升了20%.The electromagnetic signal recognition model based on deep learning has the advantages of high efficien⁃cy,high accuracy,and less manual intervention.However,it is as susceptible to adversarial examples as traditional neural network models.Studying adversarial examples is important for testing and improving the security and robustness of neural network models.In order to generate high-quality electromagnetic signal adversarial examples,this paper proposed a batch feature points Jacobian-based saliency map attack method(BP-JSMA).Compared with the traditional Jacobian saliency map attack method,BP-JSMA can generate adversarial examples faster by selecting key feature points in a batch.In addi⁃tion,according to the characteristics of electromagnetic signal data,adaptive disturbance limitation is proposed to make the generated adversarial examples more covert.Experimental results on public datasets show that compared to the Jacobian sa⁃liency map attack method,BP-JSMA improves generation speed by 11 times and concealment by 10%.Compared with tra⁃ditional fast gradient sign attack method,the attack success rate has been improved by 24%,and the concealment has been improved by 20%.

关 键 词：人工智能 深度学习 对抗样本 电磁信号识别 显著图 目标攻击 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]