적대적 공격에 강인한 적대적 학습 기반의 초분광 영상 분류기 설계

Abstract

Recently, the importance of attack and defense based on adversarial examples has been highlighted in the defense field. In particular, it is necessary to design a robust network because a black box attack can significantly deteriorate the classifier performance of the network, even if a slight change is introduced by attacking the video data. Furthermore, when the details of the network are not known, the network can be fooled through adversarial attacks. In the hyperspectral field, various classifiers have been designed using deep learning. However, these classifiers produce results that are vulnerable to adversarial attacks through the backpropagation process. This paper proposes the design of a hyperspectral classifier that is robust against adversarial attacks. The robustness of the network is demonstrated by analyzing the results of classifying hyperspectral images including various objects such as grass and objects with colors similar to that of grass. It is assumed that the paint pertains to the camouflage of a tank. The results demonstrate the significance of the hyperspectral data used in the defense field in the context of adversarial attacks. Moreover, a useful adversarial training scheme for hyperspectral classifiers is described.