Controllable face soft-biometric privacy enhancement based on attribute disentanglement

Though the widespread application of face recognition systems facilitates user authentication and identification, it may result in potential changes in user privacy concerns. This is mainly because untrustworthy service providers utilize advanced deep learning models to automatically extract users' soft-biometric attributes without user consent, thereby posing significant privacy threats. Existing attribute privacy protection methods can obscure multi-attribute while preserving identity information, but they fail to adjust the level of privacy protection, so they have limited flexibility in various scenarios. To solve this problem, this paper proposes an attribute disentanglement network (ADNet) to generate perturbed images using attribute codes from the attribute disentanglement module (ADM), thereby obfuscating arbitrary classifiers while preserving identity recognition. Specifically, the designed ADM can separate attribute-related and attribute-unrelated codes from attribute codes, realizing privacy protection during the transformation of attribute-related code and preserving unrelated information. Moreover, a control factor alpha\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document} is used to adjust the degree of transformation, and its value can be adjusted to meet various requirements. Extensive experimental results indicate that ADNet provides controllability for multi-attribute privacy protection while maintaining identity recognition utility. It meets the demands of different protection scenarios and outperforms previous soft-biometric privacy protection strategies.