Adversarial Cross-laser Attack: Effective Attack to DNNs in the Real World

Deep neural networks for image classification have been widely used to enhance user experience, but adversarial attacks continue to pose a threat to the security of deep neural networks related systems. To advance the development of more robust and secure models, research in this field is important. Laser attacks have overcome some issues of previous attacks. To enhance the laser-based attack strategy, we propose a novel physical adversarial attack method that optimizes cross-laser parameters using a Bayesian Optimization algorithm improved by contour detection technology. Our method solves the problems of traditional laser-based methods including positioning, insufficient perturbation intensity, low optimization efficiency, lack of multi-angle robustness, and optical path continuity issues. Digital and physical experiments were implemented, and our method achieved an attack success rate of up to 86.24%. Our adversarial attacks pose new challenges and requirements for artificial intelligence security.