Fully Covered Adversarial Camouflage Against Remote Sensing Detection via Physics-Driven Rendering and Pyramid Training

Uncrewed aerial vehicles (UAVs) imaging object detection systems based on deep neural networks (DNNs) are highly vulnerable to adversarial attacks. However, existing methods neither fully considered the complex physical interferences in the real world nor the multiscale and temporal observation characteristics of UAV-based detection, which makes the generated camouflage difficult to maintain effectiveness in practical tasks. To address these limitations, a fully covered robust adversarial camouflage generation framework for UAV-based detection models is introduced. First, the physically based differentiable renderer (PBDR) is designed to simulate global illumination, surface reflectance, and atmospheric scattering, thereby incorporating complex physical interference factors during camouflage training. Second, the pyramid camouflage training framework (PCTF) is introduced, which conducts hierarchical training on distance-varying inputs to preserve the multiscale adaptability of camouflage features across layers and ensure interframe consistency within layers. Furthermore, the limitations of existing evaluation indicators for UAV-based adversarial attacks are analyzed, and a novel set of indicators is proposed for evaluating physical adversarial camouflage in remote sensing. Extensive cross-domain experiments in both digital space and physical world demonstrate superior performance compared with state-of-the-art (SOTA) methods. The proposed approach achieves a temporal attack success rate (TASR) of 84.5%, see demo video at: https://youtube.com/@meyond_paa?si=foP1y4MNcC4ByZoi, compared with baselines, the average precision (AP) scores across multiple black-box detection models decrease by 18.4% on average, while the attack success rate (ASR) increases by 70.4%.