Research On Security Of Face Recognition System Based On Convolutional Neural Network

As Internet and storage technologies develop,massive data continue to emerge.Traditional statistical methods are difficult to process massive data.Machine learning algorithms based on statistical theoretical frameworks have begun to be used to parse and learn massive data.Deep learning algorithms based on convolutional neural networks are one of the most widely used machine learning algorithms,but they have fatal weaknesses.Convolutional neural networks are easily deceived by maliciously modified inputs and generate false predictions.This modified input is the adversarial example.In some tasks with high security requirements,this is a great security hazard,such as face recognition technology.Therefore,studying the problem of adversarial examples in convolutional neural networks has become an important work.Aiming at the adversarial examples in convolutional neural networks,this paper takes the face recognition system as an example to study the potential security risks of face recognition technology,and proposes mask adversarial example generation methods and corresponding defense methods.The main research is as follows:1.Aiming at the security problem in face recognition,a mask adversarial example generation method is proposed to generate a mask with disturbances,so that the face recognition model will misjudge the face wearing the disturbance mask as the target person,revealing the existence of face recognition technology safety hazards.Experiments show that the mask adversarial example mimics the situation of wearing a mask on the face,and has a certain degree of camouflage.The attack success rate is as high as 90.43%.2.Aiming at the mask adversarial example,the De Fense-EC defense method is proposed,which uses image inpainting technology to reconstruct the adversarial example into the original sample and eliminate the disturbance.Experiments on multiple data show that De Fense-EC has high quality of reconstructed images,and the defense success rate is as high as 92.32%.This paper studies the problem of adversarial examples in face recognition technology based on convolutional neural networks,and proves the harm of mask adversarial examples and the effectiveness of the De Fense-EC defense method.With the widespread application of convolutional neural networks,the problem of adversarial examples will gradually become a research hotspot.In-depth study of adversarial examples will help the further development and application of convolutional neural networks.