LPLA: The Adversarial Attack Against License Plate Recognition Systems

Neural networks are vulnerable to attacks from crafted adversarial examples. The license plate is the only sign of the vehicle; it's legally forbidden to cover or scribble on a license plate in numerous nations. The adversarial attacks against license plate recognition-based neural networks are challenging. License plate recognition is divided into two steps: localization and character recognition. In this paper, we propose a license plate location attack for the license plate detection model, which can reduce the model's prediction accuracy. Specifically, the origin-framed patch is generated by the generator with the license plate seed. Subsequently, the origin-framed patch size is adjusted based on the actual size of the license plate in images from camera equipment to create the special-frame patch. The special-frame patch is embedded around the license plate to evade license plate detection. The special-frame patch ensures the license plate remains clear and undisturbed. Many experiments show that our adversarial methods can fool license plate detection models such as Yolov5, Yolov6, Yolov7, and Faster R-CNN.