面向空间视觉目标检测的对抗攻击与防御算法  

Adversarial attack and defense algorithms towards space visual object detection

作　　者：周栋 孙光辉 吴立刚[1] ZHOU Dong;SUN Guang-hui;WU Li-gang(School of Astronautics,Harbin Institute of Technology,Harbin 150001,China)

机构地区：[1]哈尔滨工业大学航天学院,哈尔滨150001

出　　处：《控制与决策》2024年第7期2161-2168,共8页Control and Decision

基　　金：国家自然科学基金面上项目(62173107)。

摘　　要：随着航天航空技术的发展,空间目标视觉检测技术作为航天器智能在轨服务的重要技术支撑,获得了国内外研究学者的广泛关注.考虑到太空中恶劣的光照条件以及未知的动态场景,空间目标视觉检测的鲁棒性问题亟待深入研究.首先,提出一种黑盒迁移实例攻击方法,将图像识别领域的对抗样本攻击方法应用于空间目标检测任务,实现对EfficientDet目标检测模型的欺骗攻击;同时,提出一种协同防御策略,将对抗训练和SRMNet去噪器相结合,有效增强目标检测模型的鲁棒性.实验结果表明,所提出防御策略不仅能够成功抵御对抗样本攻击,还能取得高于原始空间目标检测模型的检测精度.With the development of aerospace technology,space object visual detection as the key methodology of intelligent on-orbit service of spacecraft has garnered broad concerns.Considering the extreme illumination condition and unknown scenario dynamics,the robustness problem of space object visual detection is urged to be studied in depth.This paper proposes a black-box transferred instance attack,which applies adversarial attacks in image classification domain to the task of space object visual detection.It succeeds to fool the EfficientDet model.Meanwhile,we further put forward a cooperative defense strategy that combines adversarial training with the SRMNet denoiser,which effectively enhances the robustness of the object detector.Experimental results show that this defense strategy not only resists adversarial attacks successfully,but also makes a great improvement on the accuracy of the space object detection model.

关 键 词：空间目标视觉检测 黑盒迁移实例攻击 对抗训练 SRMNet 

分 类 号：TP274.5[自动化与计算机技术—检测技术与自动化装置]