A Filtering Algorithm for Spacecraft Detection to Resist Adversarial Patch Attacks

Deep learning has been widely applied in the field of spacecraft detection. However, adversarial patches can arbitrarily manipulate image pixels within a restricted area, leading to model misclassification. Spacecraft detectors struggle to resist patch attacks caused by the occlusion of space debris or patterns carried by the spacecraft itself. Existing defense methods, while resisting patch attacks, interfere with the detection of clean images. In this article, we propose a filtering algorithm capable of resisting patch attacks. First, we trained a patch detector based on ResNet50, which determines whether an image has been subjected to patch attacks by sliding a mask. Then, we trained a patch segmentor based on U-Net, which can remove the entire patch from the image to eliminate interference. We conducted experiments on a synthetic dataset generated by Blender software and public datasets such as spacecraft pose estimation dataset (SPEED). Our experiments demonstrate that under the same attack scenario, compared with previous methods, this algorithm can improve the accuracy by 5%-30% while maintaining the same detection performance for clean images. This filtering algorithm can also achieve a defensive effect on the local spacecraft pose measurement platform.