High Similarity Controllable Face Anonymization Based on Dynamic Identity Perception

Abstract In the meta-universe scenario, with the development of personalized social networks, interactive behaviors such as uploading and sharing personal and family photos are becoming increasingly widespread. Consequently, the risk of being searched or leaking personal financial information increases. A possible solution is to use anonymized face images instead of real images in the public situations. Most of the existing face anonymization methods attempt to replace a large portion of the face image tomodify identity information. However, the resulted faces are often not similar enough to the originalfaces as seen with the naked eyes. To maintain visual coherence as much as possible while avoidingrecognition by face recognition systems, we propose to detect part of the face that is most relevant tothe identity based on saliency analysis. Furthermore, we preserve the identification of irrelevant facefeatures by re-injecting them into the regenerated face. The proposed model consists of three stages.Firstly, we employ a Dynamic Identity Perception Network (DIPNet) to detect the identity-relevantfacial region and generate a masked face with removed identity. Secondly, we apply Feature Selectionand Preservation Network (FSPNet) that extracts basic semantic attributes from the original face andalso extracts multilevel identity-irrelevant face features from the masked face, and then fuses theminto conditional feature vectors for face regeneration. Finally, a pre-trained StyleGAN2 generator isapplied to obtain a high-quality identity obscured face image. The experimental results show that theproposed method can obtain more realistic anonymized face images that retain most of the originalfacial attributes, while it can deceive face recognition system to protect privacy in the modern digitaleconomy and entertainment scenarios.