Researches On Cancelable Face Template Protection Algorithms

In recent years,with the rapid development of identity identification technology,it has been an indispensable part of application scenarios,such as financial transactions and safety protection.Traditional identification technology relying on the users' passwords is insecurity because the passwords are easily forgotten or stolen.With the development of smart devices,biometric identification has gradually become the mainstream method.Biometrics(such as faces,fingerprints,irises and so on)are the only type of physiological characteristics or behaviors that can verify the identity of an individual and have advantages of uniqueness,difficulty in stealing and forgery.However,the emergence of biological information leakage and even abuse has caused great security risks.Therefore,how to protect biometrics that are used for authentication has attracted extensive attention from researchers.Cancelable biometrics template technology is an important way to protect the original biometrics.Most technologies extract features from the original biometrics,and protect the biometrics by encrypting or transforming the features at the domain of digital image processing or feature processing,i.e.,in the digital domain.However,in the physical world,none of the techniques exploit the way to protect biometric information before registration in the system.Thus,in order to protect the user's original biological information,the concept of cancelable face templates with Hard-Template is proposed in this paper.Hard-Template(HT)is a personalized and customized image sticker that can be pasted to the user's face.The Hard-Template needs to be worn every time when the user registers and performs authentication,and it can be combined with the existing biometric template protection technology to protect the original face information better.Specifically,the work of this paper is shown as the following aspects:(1)Hard-Template based cancelable face templatesA cancelable face template algorithm based on adjustable Hard-Template in positions and appearances is proposed in this paper.First,two block-based methods are used to calculate informative areas on the face.One is to calculate the discrete entropy of the face image block.The other is to calculate the distance between the whole and an image block missed face images.Then,we use the FGSM algorithm to generate the Hard-Template.The initialization state and generation parameters of the Hard-Template are used as variables to control the appearance.When the HardTemplate is leaked,the position or appearance can be changed to form a new cancelable face template in case of the leaked information will be used for illegal activities.(2)Block-based Hard-Template cancelable face templatesIn order to ensure the HT can protect the face information,at the same time,not affect others' recognizing him because the whole HT covers large face areas,a block-based Hard-Template algorithm for cancelable face templates is proposed in this paper.First,the FGSM algorithm is used to generate a Hard-Template for a given area of the face.Then the whole Hard-Template is divided into several blocks.The appropriate replacement position of the whole Hard-Template or the position of the Hard-Template blocks could be obtained by optimizing the objective function,which is constructed by the relative entropy.Finally,the whole Hard-Template or the blocks are placed at the replaceable or selected position as a cancelable face template respectively.(3)Interpretable patterns-based cancelable face templatesIn order to improve aesthetics,a cancelable face templates algorithm based on interpretable patterns is proposed in this paper.The backbone network(HTGAN)is constructed by Generative Adversarial Networks and Deep Residual Networks.The Hard-Template is generated by the generator and combined with the original face image to obtain the target face image.Then,the original face image and the target face image are simultaneously input to the discriminator and Res Net for identification and classification.Finally,by gradually optimizing the loss function,the optimal Hard-Template is obtained that could be used for cancelable face template.