基于Grad-CAM的电磁信号对抗攻击方法  

Adversarial Attack Algorithm for Electromagnetic Signal Based on Grad-CAM

作　　者：周侠 张一然 张剑[1] ZHOU Xia;ZHANG Yiran;ZHANG Jian(Wuhan Digital Engineering Institute,Wuhan 430205)

机构地区：[1]武汉数字工程研究所,武汉430205

出　　处：《舰船电子工程》2023年第6期204-208,共5页Ship Electronic Engineering

摘　　要：随着深度神经网络在电磁信号识别上的应用越来越广泛,为提升我方智能攻击能力,使得敌方智能模型陷入瘫痪,论文将深度学习可解释方法Grad-CAM引入到对抗样本生成领域,通过生成攻击目标类别t的显著特征图,然后结合梯度下降的方式增加样本在类别t上的分类得分,直至模型将其识别为t。实验表明,显著图的对抗攻击方法能够进行针对性攻击,大幅减少无关数据点的扰动,在尽可能少地改动数据点数量的情况下就能完成对抗攻击。With the application of deep neural network in electromagnetic signal recognition becoming more and more exten⁃sive,in order to improve our intelligent attack capability and paralyze the enemy's intelligent model,this paper introduces the in-depth learning interpretable method Grad-CAM into the field of adversary sample generation,by generating the salient feature map of the attack target category t,and then increasing the classification score of the sample on category t through gradient descent until the model recognizes it as t.The experiment shows that the adversary attack method with saliency map can carry out targeted at⁃tack,greatly reduce the disturbance of irrelevant data points,and complete the adversary attack with as few changes to the number of data points as possible.

关 键 词：深度神经网络 电磁信号识别 对抗样本 显著图 

分 类 号：TN929.5[电子电信—通信与信息系统]