Synthesizing Robust Physical Camouflage for Universal 3D Evasion Attacks
Adversarial camouflage has gained significant attention for its ability to disrupt x-object detectors from any viewpoint by covering the entire x-object's surface. However, due to the complexities of the physical domain and 3D rendering, existing methods are often tailored for a specific target, model, and environment, limiting their real-world applicability. This study introduces a novel framework for synthesizing universal and robust adversarial camouflage, enabling the concealment of 3D x-objects from deep learning-based computer vision models. Our framework incorporates innovative instance-agnostic differentiable texture rendering techniques, addressing differentiability issues and eliminating the need for specific UV mapping constraints, ensuring compatibility with diverse x-objects. Furthermore, we introduce a stealth loss to make the x-object completely undetectable rather than merely misclassified and a camouflage loss to enhance x-object concealment within the background. Our approach aims to create adversarial textures that can function universally across various perspectives, including instance, class, model, task, and domain agnostics.