Research On The Security Of Social Authentication Based On Adversarial Examples Technology

In recent years,there has been an explosive growth in network-related technologies along with the popularity of 4G technology and the development of 5G technology.Social networking services(SNS)such as Facebook and Twitter are gradually being integrated into people’s lives.Social pages are rich in interactive features that prompt users to actively share a lot of personal or social information,such as personal location,personal life photos and friendships.These social networking services allow users to provide a variety of data that is used to provide better services to users.In social networks,the data that users unintentionally provide or share can lead to many security issues.In recent years a new type of social authentication system has emerged that authenticates users’ identities based on their social information.For example,in Facebook’s social authentication system’s Knowledge Authentication module.After getting the images posted in the user’s social circles,hackers target the user’s account,resulting in the risk of theft of the user’s account.This leads to an attack on the social authentication system based on social knowledge.Along with the emergence of computer arithmetic power and massive data,various deep neural networks,especially convolutional neural networks,have also been developed and obtained great results in the fields of image recognition,natural language processing,etc.In the past two years,in the field of imaging,researchers at home and abroad have found that the successful training of deep neural networks will be successful against some samples.In the past two years,in the field of image,domestic and international researchers have found that the models successfully trained on convolutional neural networks can be successfully attacked by some adversarial examples.The counter samples add some special "noise" in the image,which can lead to the misclassification of the image with "noise" in the convolutional neural network.adversarial examples can be used to defend against vulnerable situations such as knowledge validation modules.Aiming at the security problem of social accounts in social networks,this paper proposes a security mechanism based on knowledge verification accounts.This mechanism can effectively defend against attacks against convolutional neural network,thus ensuring the security of social accounts.The main contributions of this paper include:(1)The FGSM attack Algorithm and the Deep Fool attack algorithm are used to attack low-pixel images.In-depth study of the two algorithms aimed at low- pixel,small image attack actual effect.(2) The sensitive weight and the dynamic attack region are proposed,and the threshold of the sensitive weight in the case of low Pixel and small image is obtained by the experiment,and the dynamic attack region is obtained by the convolutional neural network.A new algorithm,lpixel attack,is proposed.Th range and area of attack are not interfered by human.(3)Attack on Facebook’s knowledge authentication system.The lpixel attack method is used to defend against the attack of FACEBOOK’s knowledge verification system.In this paper,the lpixel attack algorithm is studied and implemented.The lpixel attack algorithm is used to automatically generate adversarial examples from low pixel original sample images,propose sensitive weights and dynamic attack regions,and the attack success rate is over 96%.