Abstract: With the rapid development of deep learning technology, CNN has got great accuracy in SAR ship classification. However, because of the character of SAR picture and the fragility of CNN, the performance of CNN is unstable that causes hidden danger in practical?application. For CNN's insufficient robustness in SAR ship identification task, this paper makes use of adversarial example to research the adversarial robustness of SAR ship identification that represent the ability of maintaining stable input-output relation under small change. This paper use kinds of adversarial attack based on gradient, boundary, block box and so on to fool most common CNN in SAR ship identification task. Then, we use the identification result and network visualization technology to evaluate the CNN’s robustness, finding that most SAR ship identification CNN’s adversarial robustness is weak that would be easily fooled by little change. Finally, we enhance the robustness of CNN targeted based on the evaluation that behaves much better in adversarial robustness. The use of adversarial example and the above research process breaks new ground in the research for the Robustness of SAR Ship identification.