PL-attack: Projected Light-based Adversarial Attacks against Deep Neural Networks in the Physical World

The rapid development of artificial intelligence has led to the increasing application of image recognition in autonomous driving and surveillance. However, these systems are often vulnerable to adversarial attacks. In this paper, a multi-target optimized projection attack method based on image recognition is designed and implemented to evaluate and improve the security of image recognition system. By controlling the projection parameters, you can control the color, clarity, shape, and size of the projection interference, projecting it onto the object to test the recognition effect. We built a physical platform, implemented the existing attack, and then designed a multi-target optimized projection attack method to optimize the projection properties through particle swarm optimization. Tests on ResNet18 and custom fruit models have shown a 95% success rate for indoor and 90% outdoor dark attacks. This highlights the vulnerability of image recognition to physical adversarial attacks and highlights the need to enhance the system's anti-interference capabilities, thus providing insights for protecting intelligent systems.