Research On SAR Target Recognition Method Based On Deep Learning Adversarial Sample Attack

Synthetic Aperture Radar(SAR)is a kind of high-resolution imaging radar,which has the characteristics of all-day,all-weather and long-distance.It is widely used in civil and military fields so that SAR target image recognition has always been a research highlights in radar image interpretation.Compared with the traditional target recognition algorithm,the radar target recognition algorithm based on deep learning has the advantage of end-to-end feature learning,which can effectively improve the target recognition rate and become an important method of radar target recognition.However,recent studies show that the optical image recognition method based on deep learning is vulnerable to the attack adversarial examples.The existence of adversarial examples shows that deep learning method has great security risks.In SAR target image recognition,it is still an open question whether there are adversarial examples in target recognition based on deep learning algorithm.Different from optical image target recognition,the input data of radar target recognition system can be visual image in time domain or complex image in frequency domain.Based on these two data forms,this paper research the method of deep learning recognition algorithm adversarial examples attack SAR target:(1)When the input data of radar target recognition system is the time domain visual image,the target and non target attack algorithms are designed in the case of white box and black box respectively.In the case of white box,the gradient based adversarial examples generation algorithm is used to attack three deep learning algorithms for SAR target recognition.In the case of black box,the algorithm of adversarial examples generation based on decision boundary is used for targeted attack;In this paper,we use the characteristic that the adversarial examples can be migrated to attack no-target.The experiment is based on radar data.The results show that the SAR image target recognition algorithm based on deep learning has the potential risk of being vulnerable to adversarial examples attack.(2)When the input data of radar target recognition system is the complex image in frequency domain,this paper proposes an algorithm to generate the adversarial examples of complex generation adversarial network.Firstly,the traditional typical deep learning model is modified to complex form;Secondly,the complex deep learning model is used as the generator and discriminator to generate the antagonism neural network,and the discriminator is pre trained;Thirdly,the traditional real cross entropy loss function is modified to complex form,and the complex form discriminator is trained;Finally,the complex generator is used to generate the complex form of radar target adversarial examples to attack other complex deep learning models.The experiment is based on radar data.The results show that the proposed algorithm can not only attack the complex deep learning model effectively,but also has the characteristics of cross data domain attack.That is to say,when the complex adversarial examples generates the visual image in time domain,it can also attack the real deep learning model effectively no-target attack.