Digital-to-Physical Transfer of Adversarial Patches for Aerial Vehicle Detection

Jung Heum Woo, Eun-Kyu Lee · 2026

Read Digital-to-Physical Transfer of Adversarial Patches for Aerial Vehicle Detection by Jung Heum Woo, Eun-Kyu Lee (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-00159.txt.

Introduction

Woo and Lee test if fake patches that fool sky watchers on screen also work in print. Nets that spot cars from the air help track land and town growth. Printed cheat sheets could blind those nets in real life. The pair bridge screen tests and field use.

They tune patches to drop the top spot score a net gives. Rules on print range and smoothness keep the sheets real. Printed sheets are then placed on, off, and off to the side of cars. A YOLOv3 net is the test target. Off sheets win on screen, with top cuts in score. On sheets prove more tough in the field. Rain and sun tweaks in training do not help much.

Catalog record

CREATORJung Heum Woo, Eun-Kyu Lee
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.00159

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Digital-to-Physical Transfer of Adversarial Patches for Aerial Vehicle Detection Jung Heum Woo, Eun-Kyu Lee Abstract Deep neural network (DNN)-based object detectors are widely used for analyzing aerial and satellite imagery in applications such as environmental monitoring and urban analytics. Despite their strong performance, these models are known to be vulnerable to adversarial examples, and physical adversarial attacks using printable patterns pose realistic security threats. In this paper, we evaluate physical adversarial patch attacks against an aerial vehicle detector by bridging digital optimization and real-world deployment. Adversarial patches are optimized in the digital domain using a loss function that minimizes the maximum objectness score while incorporating non-printability score (NPS) and total variation (TV) constraints to ensure both printability and spatial smoothness. The optimized patches are printed and deployed in three configurations: ON, OFF, and OFF-Side. Experiments using a YOLOv3 detector show that while the OFF patch achieves the highest effectiveness in the digital domain (85.51% Average Objectness Reduction Rate (AORR)), the ON patch demonstrates superior robustness in physical environments (0.197–0.343 Objectness Score Ratio (OSR)) due to its consistent visibility. Furthermore, our results indicate that weather-based augmentation does not necessarily improve patch optimization in this domain. These findings provide critical insights into the practical vulnerabilities of aerial object detection systems. Keywords: aerial image object detection; adversarial patch; patch optimization; physical adversarial attack; aerial vehicle detection 1 Introduction The rapid growth of aerial and satellite imagery has significantly expanded the use of automated visual analysis systems in a wide range of applications, including environmental monitoring, urban planning, traffic analysis, and disaster response. To efficiently analyze such large-scale image data, deep neural network (DNN)-based object detection models have been widely adopted. Modern object detectors have demonstrated strong performance in identifying objects such as vehicles, buildings, and infrastructure in aerial imagery. Among these models, YOLOv3 has been widely used due to its high detection accuracy and real-time inference capability ref-yolov3 . Despite their strong performance, DNN-based detectors are known to be vulnerable to adversarial examples, which are carefully crafted perturbations designed to manipulate model predictions. Previous studies have shown that adversarial perturbations can significantly degrade the performance of both image classification and object detection models. While early research mainly focused on digital adversarial attacks that directly modify pixel values in images, recent work has demonstrated the feasibility of physical adversarial attacks, in which adversarial patterns are printed and placed in real-world environments to influence camera-captured inputs. However, most existing studies on physical adversarial attacks have focused on ground-based scenarios, such as autonomous driving systems or surveillance cameras. In contrast, aerial and satellite imagery present unique challenges due to large observation distances, varying viewing angles, scale variations, and environmental conditions such as illumination changes and weather effects. As a result, the effectiveness and transferability of physical adversarial attacks in aerial imagery remain relatively underexplored. Unlike ground-level photography, aerial and satellite imagery is captured through long atmospheric paths, making it susceptible to visibility degradation caused by weather conditions such as fog, haze, and fluctuating illumination. Such environmental variability can significantly undermine the effectiveness of adversarial patches when they are transferred from digital optimization to physical deployment. To address this, we incorporate weather-based augmentation into the patch optimization pipeline to simulate atmospheric effects and environmental noise, thereby investigating its impact on ensuring the robustness of adversarial attacks under diverse real-world capture conditions. Building on this motivation, this study evaluates physical adversarial patch attacks against an aerial vehicle detector by developing and analyzing a practical attack pipeline that bridges digital optimization and real-world deployment. Adversarial patches are first optimized in the digital domain and then physically printed and deployed in real environments to evaluate their impact on object detection performance. We consider three patch configurations: ON patch placed directly on the vehicle roof, OFF patch placed around the vehicle, and OFF-Side patch in which two patches are placed on both sides of the vehicle. The target detector is pre-trained on the Microsoft COCO dataset ref-coco and fine-tuned on the Cars Overhead With Context dataset ref-cowc for aerial vehicle detection. Our experimental results yield several critical insights into the physical robustness of aerial vehicle detectors. First, we found a significant discrepancy between domains: while the OFF patch configuration—which manipulates contextual information—is most effective in digital simulations (achieving up to 85.51% AORR), the ON patch demonstrates superior effectiveness and robustness in real-world environments. Second, our analysis indicates that weather-based augmentation during the optimization stage does not necessarily improve attack performance and can even hinder the generation of effective adversarial patterns in this specific domain. Finally, we demonstrate that moderate total variation (TV) regularization is essential for balancing the trade-off between physical printability and adversarial potency. These findings provide insights into the practical vulnerabilities of aerial object detection systems under real-world physical perturbations. The scientific novelty and main contributions of this paper are summarized as follows: • We present a digital-to-physical evaluation of adversarial patch attacks for aerial vehicle detection, focusing on the practical transfer gap between digital optimization and real-world aerial capture. • We compare object-level and context-level patch deployment strategies, including ON, OFF, and OFF-Side configurations, and show that the most effective strategy differs between digital and physical domains. • We analyze weather-based augmentation and TV regularization in aerial patch optimization, providing empirical evidence that environmental augmentation does not necessarily improve attack effectiveness in this setting. 2 Background and Related Work 2.1 Aerial Image Object Detection Object detection in aerial and satellite imagery has become an important research topic due to its wide range of applications, including urban analysis, environmental monitoring, and traffic understanding. Compared with natural-image object detection, aerial image detection is more challenging because target objects often appear very small, densely distributed, and visually ambiguous under varying viewpoints and imaging conditions. To address these challenges, convolutional neural network (CNN)-based detectors have been widely adopted in remote sensing tasks. Among early large-scale applications, Albert et al. ref-albert demonstrated that convolutional networks combined with satellite imagery can effectively capture large-scale urban patterns, showing the practical value of deep learning for aerial scene understanding. In addition, Ji et al. ref-ji proposed a vehicle detection framework for remote sensing imagery that integrates super-resolution with object detection to improve the detection performance of small vehicles in aerial images. These studies demonstrate that deep learning-based approaches have become the dominant paradigm for aerial object detection. Recent studies have further improved aerial and remote sensing object detection by addressing the unique characteristics of overhead imagery. Cai et al. ref-pkinet proposed PKINet to capture multi-scale local features and long-range contextual information for remote sensing detection. Huang et al. ref-mutdet investigated detection pre-training for remote sensing scenes and showed that object-level representation alignment can improve transfer performance, especially under limited annotation settings. These advances indicate that aerial object detectors are becoming increasingly specialized for scale variation, dense object layouts, and arbitrary orientations. More broadly, the success of deep neural networks in aerial imagery has motivated the development of detection systems for vehicles and other small objects observed from overhead viewpoints. In this context, aerial object detection serves as a critical foundation for many downstream tasks, and its reliability is particularly important when the detection results are used in real-world monitoring or decision-making systems. Despite their strong performance,

Read the complete text