Robust Physical Adversarial Example Generation Method Based on UViT-CycleGAN

• 1、Cyberspace Security School，Beijing University of Posts and Telecommunications，BeiJing，100876

Abstract：The rapid development of artificial intelligence, while facilitating everyday life, has also raised numerous security concerns. Physical adversarial attacks have emerged as a crucial research area revealing security issues in real-world deep learning models and promoting adversarial defense technologies. However, due to various physical constraints, physical adversarial examples face challenges in robustness. This paper addresses the issue that physical constraints in the real world can diminish the adversariality of physical adversarial examples. It proposes a robust physical adversarial example generation method based on UViT-CycleGAN, which uses UNet-ViT as the generator to refine CycleGAN. This method seeks disturbances that fit the actual physical image distribution during the process of simulating digital to physical image transformations, thereby generating more realistic and robust physical adversarial examples. Experimental results show that UViT-CycleGAN surpasses FGSM, PGD, CW, AdvGAN in terms of attack efficiency and success rate, and achieves better outcomes in enhancing the robustness of physical adversarial examples compared to current mainstream robustness enhancement frameworks EOT, RP2, and D2P.

Keywords： Robustness Physical Adversarial Examples Cycle-Consistent Adversarial Networks