Research And Application Of Face De-identification Algorithm Based On Generative Adversarial Networks

With the intelligentization and informationization of life,video surveillance,face recognition technology,social networks,etc.have provided shortcuts to privacy violations.These technologies make it possible for us to be photographed and videoed at any time,and our face information is recorded everywhere.At the same time,face images have a lot of unique personal characteristics.If these face images with private information are directly disclosed,it will lead to the leakage of personal sensitive information.Therefore,the privacy protection of face images has become a hot spot in privacy protection research.At present,the common method to achieve face privacy protection is to de-identify the face,but the current general methods can not well balance privacy protection and de-identified face data availability.In order to break through the limitations of existing methods,this paper proposes a new face de-identification algorithm.The main contributions of my article is summarized as follows:(1)In order to make the de-identified face meet the privacy protection and still have diversity,a face de-identification algorithm based on evolutionary generative adversarial networks is proposed.This method uses the idea of evolutionary algorithms in the evolutionary generative adversarial networks,and improves the evolutionary generative adversarial networks for the two technical goals of privacy protection and data availability of face de-identification.The structural similarity and distance of the the face image generated by the generator and corresponding original face are added to the performance evaluation stage of the generator in the evolutionary generative adversarial networks,and it is defined as the de-identification index,so as to eliminate the individuals with poor performance,leaving the generator with good performance,and then use The trained optimal generator generates de-identifiable faces.Through the verification of two face recognition methods and the comparison with the results of traditional methods,the results show that the generated de-identifiable face maintains the diversity of the original face set while maintaining good privacy protection performance.(2)In order to verify that the generated de-identifiable face still has data availability in an identity-independent context,and can be used in the task of face age estimation,on the basis of the first method,an age estimation method based on privacy protection is proposed.This method uses the generated de-identifiable face as a dataset,and uses the DEX age estimation method to change the age estimation regression problem into a combination of age classification and softmax expectation,and performs surface age estimation on the generated de-identifiable face.Through experimental verification,the generated de-identifiable face can be used in the age estimation scene,and can obtain a better estimation result.(3)Obtain de-identified face sets that can be used in age estimation research based on face privacy protection.The dataset is based on the Morph-II dataset.The multi-task neural network model(MTCNN)is used to detect,align,and crop the face,and use the improved evolutionary generative adversarial networks to generate an de-identifiable face image with a size of 144 * 144.In this de-identifiable face set,238 images without faces are deleted,and a total of 54896 face images of 13658 people are included.