3-D Face De-Identification With Preserving Multi-Facial Attributes: A Benchmark
Yang Merik Liu, Kevin H. M. Cheng, Marko Savić, Haoyu Chen, Zitong Yu, Guoying Zhao · IEEE Transactions on Biometrics Behavior and Identity Science · 2025
Facial analysis, such as emotion detection from facial data, is inherently vulnerable to privacy due to the rich and sensitive biometric information embedded within facial features. Privacy protection can typically be achieved by modifying face images to alter features associated with facial attributes or biometric identities. However, altering facial attributes often results in the loss of critical information necessary for facial analysis applications, while most existing approaches to modifying biometric identities have not thoroughly evaluated their impact on preserving facial attributes. Additionally, 3D information like depth and surface normal has become increasingly prominent, necessitating privacy protection methods compatible with 3D data. This paper presents a pioneering study on 2D/3D face de-identification that preserves facial attributes, i.e., facial expression, gender, ethnicity. Our method processes face images such that the original identities are unrecognizable to both humans and machines, while retaining essential facial attributes. Systematic and comprehensive experimental results using three publicly available 3D face databases with evaluations on the performance of 2D/3D face/facial attribute recognition demonstrate the effectiveness of our proposed de-identification approach with preserving facial attributes. Implementation codes are also made available to reproduce our experimental results for benchmarking and further research.