SAR Attack Method based on Heat Features of Shallow Neural Networks

Data pre-screening based on machine learning is an important preliminary step for current target reconnaissance. In the pre-recognition stage, adding noise disturbance to the target image can greatly improve the security of object. Therefore, this paper proposes a new method of Synthetic Aperture Radar (SAR) image adversarial attack that extracts heat feature maps on shallow neural networks. Generate adversarial samples under the guidance of heat feature map. The adversarial examples in this article aim to reduce the Radar Cross Section (RCS) value. Considering the feasibility of guiding physical disguise in later stages, the disturbance noise in this article is designed to be block-based. The network structure of the SAR adversarial sample generation method proposed in this paper is simple. Through testing on different databases, it was found that the algorithm in this article can significantly reduce the recall rate of automatic target recognition networks.