红外遥感图像目标识别对抗算法研究  

Object Detection Adversarial Attack for Infrared Imagery in Remote Sensing

作　　者：齐嘉豪 张宇 万鹏程 李远哲 刘星月 姚爱欢 钟平[1] Qi Jiahao;Zhang Yu;Wan Pengcheng;Li Yuanzhe;Liu Xingyue;Yao Aihuan;Zhong Ping(National Key Laboratory of Science and Technology on ATR,National University of Defense Technology,Changsha 410073,China;School of Computer Science and Technology,Anhui University of Technology,Maanshan 243032,China)

机构地区：[1]国防科技大学ATR重点实验室,长沙410073 [2]安徽工业大学计算机科学与技术学院,安徽马鞍山243032

出　　处：《航空兵器》2022年第3期47-53,共7页Aero Weaponry

基　　金：国家自然科学基金项目(61971428)。

摘　　要：针对现有识别对抗算法对小尺度目标攻击效果差、对抗样本中存在大量无意义扰动、扰动生成效率低等问题,以红外遥感为应用背景,基于对抗生成攻击理论提出一种具有较强泛化性的目标识别对抗算法。算法引入空洞卷积和注意力机制构造多通道变尺度扰动生成网络以克服红外遥感图像存在的小目标问题;同时,基于检测热力图设计滤波器对生成扰动信息进行筛选,实现无意义扰动消除;最后,以第三届“空天杯”全国创新创意大赛复赛所公布数据集为例进行实验分析。与次最优攻击算法相比,本文所提算法的平均攻击成功率提升了0.313,同时将生成对抗样本的平均耗时降低了57.409 s;此外,利用生成的对抗样本去迁移攻击其他类型的检测器,使得YOLOv3检测器、YOLOv5检测器和Faster-RCNN检测器的平均检测精度分别下降了0.032,0.287和0.09。实验结果表明,本文算法在对抗样本物理可实现性、迁移性和生成速度方面都具有显著优势。Aiming at the problems of poor effect of existing adversarial attack for object detection algorithms on small-scale target attack,a large number of meaningless disturbances in adversarial samples and low disturbance genera-tion efficiency,taking infrared remote sensing as the application background,a object detection adversarial attack algorithm with strong generalization is proposed based on the adversarial generation attack theory.The dilated convolution and attention mechanism are employed to construct multi-channel and various scale disturbance generation network to overcome the problem of small targets in infrared remote sensing images.Meanwhile,a filter is designed based on the heat map of detection result to filter the generated disturbance information and eliminate the meaningless disturbance.Finally,the data set published in the third"Aerospace Cup"national innovation and creativity competition is taken as an example for experimental analysis.Compared with the suboptimal attack algorithms,the average attack success rate of the proposed algorithm is increased by 0.313,and the average time of generating adversarial samples is reduced 57.409 s.In addition,using the generated adversarial samples to transfer and attack other types of detectors,the average detection accuracy of YOLOv3 detector,YOLOv5 detector and Faster-RCNN detector is reduced by 0.032,0.287 and 0.09 respectively.Experimental results show that the proposed algorithm has significant advantages in the physical realizability,transferability and generation speed of adversarial samples.

关 键 词：红外图像 识别对抗 小尺度目标 对抗生成攻击 物理可实现性 迁移性 攻击算法 

分 类 号：TJ760[兵器科学与技术—武器系统与运用工程] TN976[电子电信—信号与信息处理]