Multi-texture Fusion Attack: A Robust Adversarial Camouflage in Physical World

Adversarial camouflage, as a physical domain attack method to object detection, possesses the advantage of multi-viewpoint attacks. However, existing adversarial camouflage methods still suffer from the issue of unstable adversarial effects under different angles. To address this problem, we propose a novel framework called the Multi-texture Fusion Adversarial camouflage generation framework (MFA) to achieve robust adversarial attacks against detection models. Specifically, we improve the original neural renderer by introducing the neural fusion renderer that combines neural renderer with texture fusion module. The texture fusion module utilizes differentiable optimization to generate textures that fuse adversarial information from various textures. To further enhance robustness, we employ a series of composite transformations to fit multiple scenarios in the physical world. Through extensive experiments, we display how our proposed method successfully addresses the instability of camouflage textures under different angles. MFA consistently outperforms existing methods in both simulated and real-world environments.