Research On Reversible Face Anonymization Based On Joint Modeling Of Identity And Attribute Features

In recent years,the massive use of face images not only has greatly facilitated people’s daily life and work,but also has directly or indirectly led to a large number of privacy disclosure.Therefore,face image privacy protection has received wide attention.As generative adversarial network(GAN)has shown excellent potential in synthesizing real images,they provides new ideas for solving the problem of face privacy leakage.Nowadays,GAN is widely used for privacy protection by anonymization and has shown broad development prospects.However,existing methods still have the following problems to be addressed: 1)it is difficult to maximize the availability of facial attributes while protecting privacy;2)it is difficult to minimize loss reversible modeling of privacy protection results.For the first problem,thesis explores a face anonymization method based on multiscale feature analysis.First,we train a face synthesis network based on a self-supervised learning method to translate the input data to a realistic face image.Then,we use the kanonymity method to process the multi-scale features to protect the identity information.Finally,the anonymized multi-scale features are fed into our face synthesis network to generate the final anonymized face image.Relying on the generalization ability of the self-supervised model,the proposed method can directly help to control the userdefined non-identity attributes of the image,which would benefit the performance of attribute preservation.For the second problem,thesis explores a reversible face anonymization method based on multi-condition joint control.First,we train a multi condition control model based on a partial pre-training model.Then,we process the identity features by using a feature deflection method based on identity space sampling.Finally,we introduce a feature steganography module to achieve reversible anonymization.On top of the theoretical lossless information of reversible networks,we introduce a feature steganography module to support image recovery,and we also introduce a background stitching module to reduce the background information loss of the recovered images.This ensures the reversibility by minimizing the information loss from both the theoretical and model viewpoints.To verify the effectiveness of the proposed method,thesis conducts experiments on several public datasets.By comparing with existing methods,we find that the proposed method performs better in several evaluation metrics.At the same time,by comparing the results of our two studies,we find that our second study can further improve the privacy protection performance on anonymization and attribute preservation.Finally,we design and develop a face privacy protection system based on the studies methods.