Research On Adversarial Attack Of The Deep Learning-based Sar Target Recognition Network

Synthetic Aperture Radar(SAR)can detect targets on the earth’s surface in real-time without the influence of natural environment and human interference,and has excellent earth observation capability.SAR is widely used in military intelligence survey and intelligence collection.Among them,SAR image target recognition technology is the current hot issue in target recognition and remote sensing image interpretation,which is an important session in SAR image interpretation and has important value and application prospects.In recent years,deep learning has attracted more and more attention in the field of SAR image interpretation and target recognition with its remarkable learning ability and classification capability.Compared with the traditional SAR image target recognition methods,this method is more universal and its recognition effect has been greatly improved.However,the security of deep learning-based SAR target recognition models is still an issue to be explored.For deep learning networks,due to their end-to-end data-driven characteristics,their prediction results are easily affected by some small perturbations with deliberate modifications.In this project,for the deep learning SAR target recognition network model,in order to address the threat of malicious modification perturbations to its security,we investigate the formation of malicious modification perturbations and proposes a research on adversarial attack of the deep learning-based SAR target recognition network.On the one hand,this research can measure the decision edge of the SAR target recognition network and enhance the interpretability of the network as a way to promote the study of the reliability of the network and thus achieve the secure deployment of the recognition system.On the other hand,the SAR adversarial samples generated by this research can be used as supplementary dependent data for SAR target identification networks,which can provide support in terms of data release to own side and defensive network construction for own side.The main work of this paper is as follows.On the other hand,the SAR adversarial samples generated by this research can be used as supplementary data for SAR target identification networks,which can provide support in data release to own side and defensive network construction for own side.Aiming at the drawback that the deep learning network is easily affected by maliciously modified small perturbations,we fully investigate the existing mainstream adversarial attack algorithms,apply them to the SAR target recognition network,and provide the strengths and weaknesses of them.Next,based on the shortcomings of existing algorithms,three criteria(high attack success rate,high misclassification confidence and low perturbation range)are proposed for the adversarial attack algorithm of SAR target recognition network with the goal of better fitting the characteristics of SAR images.Based on this,we make adaptations to the adversarial attack algorithm and proposes a high-performance SAR target recognition network adversarial attack model.Finally,based on the physical realizability of the adversarial attack,this project optimizes the SAR target recognition network adversarial attack model from the perspective of practical application deployment of SAR target recognition network adversarial attack,and proposes a target region-based SAR adversarial attack method.The adversarial attack method circles the perturbation of SAR adversarial sample to the detection target area.According to the perturbation law of the adversarial sample and the imaging characteristics of SAR image,some strong scattering objects or absorbing materials can be added to the detection target in the physical world to realize the physical camouflage of the detection target,so as to realize the physically realizable SAR target recognition network adversarial attack.