Title: Translating Digital Adversarial Patches to Physical Space for Aerial Vehicle Detection

Abstract:

Deep neural network (DNN) object detectors are extensively employed to process aerial and satellite imagery, supporting critical applications like urban analytics and environmental monitoring. However, these systems are susceptible to adversarial examples, with physical attacks utilizing printable patterns presenting tangible security risks. This study investigates the efficacy of physical adversarial patch attacks against aerial vehicle detectors by connecting digital optimization processes with real-world implementation.

We generate adversarial patches through digital optimization, employing a loss function designed to suppress maximum objectness scores. To guarantee that the patches are both printable and spatially smooth, we integrate non-printability score (NPS) and total variation (TV) constraints into the optimization framework. The resulting patches are then printed and tested across three distinct deployment scenarios: ON, OFF, and OFF-Side.

Our experiments, conducted using a YOLOv3 detector, reveal a divergence between digital performance and physical robustness. While the OFF configuration yielded the highest efficacy in the digital environment—achieving an Average Objectness Reduction Rate (AORR) of 85.51%—the ON patch proved more resilient in physical settings. The ON patch maintained an Objectness Score Ratio (OSR) between 0.197 and 0.343, attributed to its consistent visibility. Additionally, the study found that weather-based augmentation did not enhance patch optimization within this specific domain. These outcomes offer significant insights into the practical security vulnerabilities inherent in aerial object detection systems.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC