Research On Adversarial Sample Generation Method For Remote Sensing Image Classification Network

Deep neural network is widely used in remote sensing image classification tasks because of its high accuracy.However,current research shows that deep learning models are threatened by adversarial sample in remote sensing classification tasks,and their robustness drops sharply in the face of adversarial sample.Studying adversarial sample generation methods can understand the risks faced by the network in advance and accurately evaluate the robustness of its robustness.So far,adversarial sample generation methods mainly focus on natural image classification networks,while research on adversarial sample generation methods for remote sensing image classification networks is still in its infancy.Remote sensing images have different requirements for the generation of adversarial sample because of their characteristic properties different from natural images.Aiming at the remote sensing image classification network,this paper studies the method of remote sensing image adversarial sample generation from two directions:single-mode remote sensing image classification network and multi-mode remote sensing image classification network,according to the characteristics of remote sensing image data and classification model under different modes,which provides the basis for the optimization and robustness evaluation of remote sensing image classification network,so as to promote the research of high robustness remote sensing image classification network.The main work of this paper is as follows:1.Aiming at the poor performance of single-mode remote sensing image classification network in capturing the features of samples located near the decision boundary,taking hyperspectral image classification network as the research object,a hyperspectral image adversarial sample generation method based on TargetFool is proposed.In this method,the most confusing samples and the most difficult to distinguish classes are selected as the base samples and target classes,and the adversarial samples with small disturbance and strong attack are constructed according to the shortest distance between the sample and the classification decision boundary.Furthermore,the generated adversarial samples are added to the training set to iteratively complete the network training and the generation of adversarial samples,so that the network can generate adversarial samples according to the real-time classification situation and adjust the decision boundary to improve the robustness of the hyperspectral image classification network.2.Aiming at the task of multimodal remote sensing image classification,the vulnerability of multimodal remote sensing image classification network is studied for the first time,and a multimodal adversarial sample generation method based on multi-feature collaborative adversarial attack is proposed.Two disturbance generators and two modal discriminators are designed,and a cooperative learning regular term is introduced in the process of anti-disturbance optimization,which makes the disturbances of the two modes restrict and promote each other in the learning process.So that the generated multimodal adversarial samples can maintain high attack performance under the minimum disturbance intensity.3.Combined with the adversarial algorithm proposed in this paper,a remote sensing image classification network adversarial robustness evaluation system is designed.The system makes the experimental process transparent,the experimental results intuitive,and the experimental effect display intelligent by loading the pre-stored adversarial samples and pre-trained network model parameters.In the adversarial robustness evaluation,not only the visual classification effect picture is displayed,but also the image information and the corresponding robustness evaluation value are displayed,which aims to evaluate the adversarial robustness of remote sensing image classification network in a more convenient way.In this paper,the effectiveness of the proposed method is verified on hyperspectral datasetsPavia University(PaviaU),University of Houston(HoustonU2013)and multimodal datasetsISPRSVaihingen2D dataset and ISPRSPotsdam2D dataset.Experimental results show that the proposed method can obtain small disturbance and strong attack samples for both single-mode remote sensing image classification network and multi-modal remote sensing image classification network.It can evaluate the robustness of remote sensing image classification network more accurately,and the research results have important theoretical and practical significance.