Research On Face Information Protection Method Based On Generative Adversarial Idea

In recent years,with the acceleration of social information process,identity authentication as the first barrier of information security system,has become more and more important.Traditional identity authentication techniques such as passwords are not secure because they can be easily forgotten or stolen.Biometrics are inherent to users,which are unique and not easy to be stolen and forged.Therefore,biometrics identification has gradually become the mainstream way of identity authentication.Among them,face recognition has been widely used because of its noncontact and easy acceptance.However,with the development of face forgery technology and the emergence of face information leakage problems,face recognition is no longer secure.In addition,the face is unique and irrevocable.Once it is forged or leaked,it will be irretrievable,and may bring unpredictable risks to the user’s personal safety and property rights.Therefore,how to protect the original face information while using face recognition technology has become an important issue.The existing face information protection technologies mainly protect the original face information by encrypting or transforming the extracted feature template after extracting the original face feature.These technologies are carried out in the digital domain,and the complete original face information needs to be obtained in the face registration stage.Thus,there is still a security risk that the original face information is leaked.Although the face anonymization technology anonymizes the face identity before extracting the original face features,the anonymized faces cannot be used for face recognition tasks.Based on the above considerations,this thesis proposes to use face accessories to protect face information,under the premise of ensuring face recognition performance,to achieve the purpose of protecting the user’s most original face information.Face accessories are masks,glasses or headbands with specific information designed according to the user’s face information in the physical domain,and images with specific information in the digital domain.When users register face recognition system and perform identity recognition,wearing face accessories can block part of the original face information on the one hand,and realize the revocability of the face on the other hand,so as to protect the original face information.At the same time,face accessories can be combined with the existing face information protection technologies to achieve multiple protection of face information.The specific works of this thesis are as follows:(1)Face information protection algorithm based on face accessories generated by GANThis thesis proposes a face information protection method based on face accessories generated by GAN.Firstly,the discrete entropy is used to calculate the information in different regions of the face,and the regions with more information are selected as the position of wearing face accessories.Secondly,this thesis combines DCGAN and Arcface to design a network MaskGAN for generating face accessories.The network can combine the original face information to generate different face accessories in different regions of the face.In the physical domain,the generated accessories can be printed and pasted to the specified position,and in the digital domain,the accessory pattern can be mapped to the specified position,so that the face wearing the accessories and the original face are no longer recognized as the same identity.Finally,the face with accessories are used to replace the original face for the face recognition tasks.Experimental results on the LFW dataset,the LFW-mask dataset constructed in this thesis and the mmsys dataset collected in the physical domain demonstrate that the proposed method can effectively protect the original face information under the premise of ensuring the recognition performance.(2)Face information protection scheme based on the combination of multiple accessories generated by GANThis thesis proposes a face information protection scheme based on the combination of multiple accessories generated by GAN.The proposed scheme uses two different accessories to protect the original face to further improve the effectiveness of face accessories for face information protection.Firstly,Mask-GAN is used to generate different accessories in different regions of the face.Secondly,this scheme combines two different accessories,including three different combinations of wearing a mask and glasses at the same time,wearing a mask and a headband at the same time,as well as wearing glasses and a headband at the same time.Finally,this thesis tests the face recognition accuracy under different combinations of accessories.Experimental results demonstrate that the scheme can protect the original face information more effectively.(3)Face information protection method based on face accessories with meaningful patternsThis thesis proposes a face information protection method based on face accessories with meaningful patterns to improve the aesthetics of face accessories.Firstly,a meaningful image is used as a reference image,and the image is forged as a face accessory image through the generator and discriminator.Secondly,the face accessory is combined with the original face image to obtain the target face image.Then the original face image and the target face image are input into the face classifier to calculate the loss between them.Finally,by gradually optimizing the loss function,the face wearing meaningful pattern accessories and the original face are no longer recognized as the same identity,and the optimized face accessories with meaningful patterns are obtained.Experimental results on the LFW dataset,the LFW-mm dataset constructed in this thesis and the mmsys dataset collected in the physical domain demonstrate that face accessories with meaningful patterns can effectively protect the original face information.