Design And Implementation Of Adversarial Examples Generation And Defense Method For License Plate Recognition

With the development of artificial intelligence technology,image recognition technology based on deep neural network has been applied in many realistic tasks,such as license plate recognition task.However,deep neural network is easy to be attacked by carefully designed adversarial examples,which affects the reliability of license plate recognition system using neural network and causes security problems.When applying existing adversarial examples generation techniques to license plate recognition tasks,on the one hand it is necessary to add human-eye imperceptible adversarial examples to the text region of the license plate,which can not deploy in the physical world;on the other hand can also be to add local-based adversarial examples to the text region,but which can easily be judged as license plate obscuring.Therefoe we propose two-stage adversarial examples training method to accomplish the task of attacking license plate recognition systems.we also investigate the use of the adversarial examples defense training method to defend against the adversarial attack.The following work is accomplished in this paper:(1)We proposed a two-stage adversarial examples generation method for license plate recognition system,that realizes the attack disturbance effect under the condition that the adversarial examples does not cover the text area of license plate image.For two different attack tasks,we designed and implemented the Yolo detection model bounding box expanded adversarial examples generation method based on NMS information and the CRNN half-targeted adversarial examples generation method based on weak character.Based on CCPD dataset subset,the proposed method is used to attack the license plate recognition system based on Yolov4 and CRNN.The results showed that the proposed method can significantly reduce the accuracy of model recognition.(2)We studied the method of adversarial examples defense training according to our generation method of adversarial examples,that can enhance the ability of extracting robust features of the model.This method can effectively defend against adversarial examples attacks.(3)We designed and implemented the generation and defense system of adversarial examples for vehicle recognition.Through this system,users can generate adversarial examples and conduct effect test.Users can customize management model and dataset,customize data to defense train and generate defense model.