Learning Robust and Repeatable Physical Camouflage for Aerial Object Detectors

Physical attacks are gaining recognition as essential tools for evaluating the security of AI systems in real-world applications. To implement these attacks, adversarial camouflage is meticulously crafted to envelop the entire surface of the target object. Existing attack methods either rely on fixed-view datasets to apply camouflage to specific surfaces of the target or require a complete 3D model of the target to achieve full coverage. However, the former approach limits the effectiveness of the camouflage across multiple viewpoints, while the latter creates a strong dependency between the camouflage and the target’s surface, making it difficult to generate easily applicable, image-like camouflage patterns.To generate reproducible and effective camouflage, we segment the target into a grid and select regions based on the detection model’s attribution map for the current image. Camouflage is then randomly applied to these grid sections, ensuring that any part effectively obscures various structures of the target. To enhance effectiveness in the physical domain and within the environment, we performed a primary color analysis of the background to identify representative environmental colors, constraining the color space of the camouflage to match these colors and improving its transparency and blending capability. By combining these perturbations, we developed a viewpoint-independent camouflage. Extensive testing on the VisDrone and DAS datasets demonstrates that our method achieves the highest attack success rates (ASR), with a maximum detection ASR of 66.91% on the VisDrone dataset, and a maximum detection ASR of 80.16% on the DAS dataset. We achieved a maximum ASR of 79.47% in the physical domain as well.