Defense against Adversarial Patch Attacks for Aerial Image Semantic Segmentation by Robust Feature Extraction

Deep learning (DL) models have recently been widely used in UAV aerial image semantic segmentation tasks and achieved excellent performance. However, DL are vulnerable to adversarial examples, which bring significant security risks safety-critical systems. Existing research mainly focuses on solving digital attacks for segmentation, but patches with physical attack attributes more threatening than attacks. In this article, we systematically evaluate the threat of task first time. To defend against patch obtain accurate results, construct a novel robust feature extraction network (RFENet). Based characteristics images patches, RFENet designs limited receptive field mechanism (LRFM), spatial enhancement module (SSEM), boundary perception (BFPM) global correlation encoder (GCEM), respectively, solve from model architecture design level. We discover that features, shape features contained can significantly enhance robustness Extensive experiments three benchmark datasets demonstrate proposed has strong resistance compared existing state-of-the-art methods.