SAR图像稀疏对抗攻击

Sparse Adversarial Attack of SAR Image

基于深度学习的图像解译技术在多个领域都取得了巨大成功,在合成孔径雷达(Synthetic Aperture Radar,SAR)图像分类、检测、分割等问题中也逐渐开始广泛应用。现有的SAR图像分类深度学习模型由于训练数据集样本量较小易过拟合,样本的微小改变易导致模型分类错误,产生对抗攻击现象。针对上述问题,本文从攻击方法、攻击结果和攻击目标三方面说明了SAR图像对抗攻击存在的问题和挑战。本文聚焦SAR图像的稀疏性,具体阐述了稀疏攻击提出背景和SAR图像中稀疏性的表现形式,并就常见稀疏攻击方法进行分析总结。文章在MSTAR数据集上验证了现有的稀疏攻击方法的有效性,分析了算法计算效率和成功率、耗时等指标,并对SAR图像分类稀疏对抗攻击方法进行展望。

Image interpretation technology based on deep learning has achieved great success in many fields,and has gradually begun to be widely used in synthetic aperture radar image classification,detection,and segmentation problems.Existing SAR image classification deep learning models are prone to overfitting due to the small sample size of the training dataset,and small changes in samples can easily lead to model classification errors,which calls adversarial attack phenomena.In response to the above problems,this article explains the problems and challenges of SAR image adversarial attacks from three aspects:attack method,attack result and attack target.This article focuses on the sparsity of SAR images,specifically expounds the background of sparsity attacks and the manifestations of sparsity in SAR images,analyzes and summarizes common sparsity attack methods.The article verifies the effectiveness of the existing sparse attack methods on the MSTAR dataset,analyzes the calculation efficiency,success rate,time-consuming and other indicators of the algorithm,and prospects the SAR image classification sparse attack method.