Adversarial Attacks And Defenses For Remote Sensing Image Classification Based On CNN Models

Deep convolutional neural network(DCNN)models have made breakthrough progress in remote sensing image classification,target detection and recognition,image segmentation and some other tasks.However,due to the vulnerability of deep neural network to adversarial examples,the application of deep learning in remote sensing has great security risks.Aiming at the remote sensing image classification task,combined with the application background of military remote sensing reconnaissance,this paper carries out relevant research on adversarial attacks and defenses.The major research of the thesis is summarized as follows:(1)In the field of adversarial attacks on multi-source remote sensing images classification,this thesis demonstrates the vulnerability of CNN through the analysis of the input and the state space and the geometric analysis of adversarial attack.Nine CNN models are tested on three kinds of data sets using nine adversarial attack methods,which shows the general vulnerability of CNN models and verifies the great threat of adversarial examples to military remote sensing reconnaissance.In addition,aiming at the problem of uncoupling adversarial noise of multi-source remote sensing images,an adversarial attack algorithm based on sparse differential coevolution is proposed,which implements the adversarial attack on multi-source remote sensing image classification for the first time.Experiments show the effectiveness of the algorithm.(2)In the field of adversarial defenses on remote sensing image classification,this thesis proposed Unsupervised Adversarial Contrastive Learning(UACL)to solve the problem that the labeled data is relatively scarce in remote sensing field,which is difficult to meet the training needs of supervised adversarial defenses,and the supervised defenses seriously reduce the standard accuracy.UACL forces unlabeled data and its unsupervised adversarial examples to obtain similar deep features through the contrastive learning of Siamese networks to enhance the robustness.Experiments show that UACL can enhance the robustness of the model while maintaining a high standard accuracy,and can be combined with the supervised defense methods to further improve the performance of CNN models.(3)In the field of adversarial evaluation on remote sensing image classification,this paper constructs an evaluation system for adversarial robustness,aiming at the problems of simple evaluation index and incomplete evaluation.The evaluation system consists of 12 metrics,including 4 data-oriented metrics to evaluate the imperceptibility of adversarial perturbation in computer vision and human vision and the neuron coverage of the adversarial examples,and 8 model-oriented metrics to evaluate the structure,behavior and activation of CNN models.Experiments on two datasets show that this system can evaluate the adversarial robustness comprehensively and provide suggestions for further enhancing the adversarial robustness of CNN models.