Abstract: Existing adversarial attack methods are limited to real-valued convolutional neural networks. To address this limitation, this paper proposes a complex adversarial sample generation method based on generative adversarial networks. Firstly, a complex model for generating effective adversarial samples is designed by introducing complex computation modules. Secondly, a pre-trained complex network is used as the discriminator in adversarial training, with a residual neural network serving as the basic framework, to enhance the attack capability of adversarial samples. Finally, transferable adversarial attacks are achieved through substitute models, resulting in higher attack success rates. Experimental results demonstrate that the success rates of the proposed method in targeted and untargeted attack tasks reach 76.338% and 87.841%, respectively, with higher transferability and closer resemblance between adversarial and original clean samples. By extending adversarial attacks to complex neural networks, this method preserves synthetic aperture radar (SAR) target information and accuracy, providing a reference solution for the security and robustness of practical synthetic aperture radar automatic target recognition (SAR-ATR) systems.

Key words: generative adversarial networks, adversarial examples, synthetic aperture radar automatic target recognition (SAR-ATR), plural convolutional neural networks, targeted attack, untargeted attack