A Robust Adversarial Patch Attack Against UAV ‐Based Aerial Object Detection

Aerial object detection systems based on deep neural networks (DNNs) are vulnerable to adversarial patch attacks. Although existing aerial patch studies have considered real‐world deployment factors, these factors are often modeled in a fragmented manner or evaluated under non‐unified protocols. Building upon prior aerial adversarial patch research, this paper formulates UAV‐based physical patch attack as a unified transformation‐aware optimization problem, in which deployment variations and layout‐dependent patch placement are jointly considered to improve robustness under diverse aerial‐view conditions. A multi‐component loss function is introduced to minimize a differentiable detection‐preserving surrogate score while jointly regularizing chromatic saliency, local smoothness, and physical printability, thereby improving attack effectiveness and physical realizability. Cross‐architecture transfer‐based gray‐box evaluation reduces the average mAP@0.5 to 0.189 on heterogeneous target detectors. Lab‐scale physical experiments further indicate the feasibility of the proposed method under controlled physical conditions, highlighting potential security vulnerabilities in UAV‐based object detection systems.