| Under the unprecedented development and application of the Internet face image big data,the artificial intelligence technology has been widely used to improve people’s daily life and work efficiency.However,since face images contain a lot of personal information,their collection and application may pose potential security risks,and may lead to the problem of privacy disclosure.Existing works reveal that face anonymity is an effective method to protect the privacy of face images,and many potential methods were proposed,especially for the methods based on differential privacy.However,the most recent works still suffer from the following drawbacks: 1)the feature-based differential privacy protection method may harm the image quality and data utility;2)the identity modification based methods may easily lead to identity invasion.To address these problems,this paper carries out the following research:To solve the first problem,a geometric modeling based face anonymization model is proposed.A conditional generation model composed of residual elements is designed based on the generative adversarial network to simultaneously protect the identity and preserve the data utility.Firstly,the original image is preprocessed to obtain its geometric structure which is further anonymized by using the Laplace mechanism based differential privacy.Then,a style encoder is used to extract the style feature information.Finally,a new anonymous face image is synthesized by the feature fusion.The proposed model is can not only flexibly control the facial geometric structure and its appearance style,but also preserve the posture of the original face.The experimental results show that the proposed method has exhibited good privacy protection performance.Also,it has exhibited some progresses on improving the performances of visual quality and attribute preservation.To solve the second problem,an identity feature deactivation based anonymization model is proposed by taking the output of a new designed CAM-based identity feature deactivation method,the geometric modeling based anonymization method and the style encoder as input and output high-quality anonymous face images with complete background.The proposed method can improve the data utility by preserving the original style and posture.Besides,it can also improve the diversity of the results by flexibly controlling the input conditions of the generative network.The experimental analysis on multiple public data sets show that the model can not only achieve a high level of identity anonymization performance but also exhibit better identity antiintrusion and utility preservation performances.Based on the B/S architecture,a controllable and diversified face privacy protection system is developed to meet users’ different demands for privacy protection.The anonymous algorithm model proposed in this paper is integrated and applied to realize the purpose of protecting the privacy of face images.The system can generate diverse anonymous results by choosing different anonymization method and parameters according to the requirements of the system users. |
