Natural Physical Adversarial Attack Method for UAV Visual Detection System

Adversarial attacks on UAV visual recognition systems present critical security challenges, particularly in physical environments. Existing physical adversarial attack methods face dual limitations of ineffective long-range targeting and conspicuous adversarial patterns. We propose a natural adversarial patch optimized for multi-scale, multi-target UAV attacks through three key innovations: (1) a unified optimization framework combining attack efficacy and visual similarity constraints, (2) physical robustness enhancements against environmental variables, (3) the proposed approach is validated through digital and physical experiments. Against YOLO detectors (v3/v5/v7), our method achieves digital attack success rates of 72.6%, 77.6%, and 75.0% respectively. Physical-world evaluations under varying altitudes ($20-100 \mathrm{~m}$) demonstrate average success rates of 52.2% (YOLOv3), 53.0% (YOLOv5), and 63.4% (YOLOv7) for adversarial patch implementation.