Adversarial attacks on deep-learning-based SAR image target recognition

Abstract

Synthetic aperture radar (SAR) image target recognition has consistently been a research hotspot in the field of radar image interpretation. Compared with traditional target recognition algorithms, SAR target recognition algorithms based on deep learning offer end-to-end feature learning, which can effectively improve the target recognition rate, making them an important method for radar target recognition. However, recent research shows that optical image recognition methods based on deep learning are vulnerable to adversarial examples. In SAR image target recognition, whether adversarial examples exist for deep learning algorithms is still an open question. This paper uses three mainstream algorithms to generate adversarial examples to attack three classical deep learning algorithms for SAR image target recognition. The experiments involve publicly real SAR images for white-box and black-box attacks. The results show that SAR target recognition algorithms based on deep learning are potentially vulnerable to adversarial examples.