Generative Adversarial Patch Attacks for Aerial Object Detection

Adversarial attacks on object detection are typically categorized into two types: global perturbations, which apply small noise across the entire image, and adversarial patches, which introduce strong localized perturbations. Global perturbations often achieve high success rates but lack controllability and task-specific focus, while adversarial patches provide more targeted and efficient attacks. In this work, we propose a patch-based attack method using a pretrained GAN generator combined with a tailored loss function to craft adversarial patches for aerial image detection. Experiments on the MAR20 dataset with YOLOv5 confirm the effectiveness of the proposed method.