Global Feature Attention Network: Addressing the Threat of Adversarial Attack for Aerial Image Semantic Segmentation

Abstract

Aerial Image Semantic segmentation based on convolution neural networks (CNNs) has made significant process in recent years. Nevertheless, their vulnerability to adversarial example attacks could not be neglected. Existing studies typically focus on adversarial attacks for image classification, ignoring the negative effect of adversarial examples on semantic segmentation. In this article, we systematically assess and verify the influence of adversarial attacks on aerial image semantic segmentation. Meanwhile, based on the robust characteristics of global features, we construct a novel global feature attention network (GFANet) for aerial image semantic segmentation to solve the threat of adversarial attacks. GFANet uses the global context encoder (GCE) to obtain the context dependencies of global features, introduces the global coordinate attention mechanism (GCAM) to enhance the global feature representation to suppress adversarial noise, and the feature consistency alignment (FCA) is used for feature calibration. In addition, we construct a universal adversarial training strategy to improve the robustness of the semantic segmentation model against adversarial example attacks. Extensive experiments on three aerial image datasets demonstrate that GFANet is more robust against adversarial attacks than existing state-of-the-art semantic segmentation models.