Abstract: Deep neural networks have demonstrated remarkable performance in various SAR intelligence interpretation task scenarios. However, it has been proved that they are vulnerable to attack from adversarial examples. For the object detection task, there have been numerous adversarial attack methods based on optical images, which cannot be implemented on SAR images directly resulting from the difference in imaging mechanism between SAR and optical images. To address the need for adversarial attacks on intelligent SAR object detection, we propose a novel method for generating adversarial patches, which performs adaptive optimization for the location and size of the patch, fuses image′s background grayscale information to enhance the patch iteration process and improves the loss function based on the specific object detection task type and target model structure. Experimental results using the SAR-Ship-Dataset demonstrate that the improved adversarial examples achieve more ideal attack effects by reducing the mAP50 of the model from 98.3% to 69.82%, mAP95 from 73.8% to 32.96%, and increasing the attack rate from 16.93% to 33.53%.