Research On Security Of License Plate Recognition System Based On Convolutional Neural Networks

Recent studies have shown that deep learning is vulnerable to attacks from adversarial examples.This presents security challenges for safetycritical scenarios,such as license plate recognition system.In this paper,the license plate recognition system is taken as the research background,and its security is studied.The main work of this paper is as follows:(1)We propose a spot attacking algorithm.By fixing the disturbance amount(that is,the color to be attacked)and taking the disturbance position as the optimization variable,we search for the best disturbance position in the license plate image,which makes the generated adversarial examples deceive the license plate character classifier based on CNN.For this reason,we put forward a method of generating "spot" adversarial examples based on genetic algorithm,which has the characteristics of no need of CNN internal parameters and finding the global optimal or approximate global optimal solution.Experiments show that the "dirty" adversarial examples generated by our proposed algorithm have high camouflage,and the highest attack success rate can reach more than 93%.(2)We propose a defense method by auxiliary GAN to against spot attacking.The generator of auxiliary GAN is a self-encoder similar to U-net.The goal of the generator is not only to repair the input adversarial examples into normal examples,but also to require the generated examples to be correctly identified by the target model.Therefore,we added the loss function similar to the C&W attacking algorithm to the loss function of the previous generator,but unlike the attacking,our loss is to classify it correctly.The experimental results show that the auxiliary GAN repair algorithm can perfectly repair the adversarial examples into the same pictures as the normal examples,and can be correctly classified by the target model with the highest accuracy of 100%.In this paper,the security of license plate recognition system based on convolutional neural network is deeply studied,and the attacking algorithm in the form of spot and its corresponding auxiliary GAN repair method are put forward,which provides a certain foundation for further research on the security of license plate recognition system.