IdentityDP: Differential Private Identification Protection for Face Images

• DOI: 10.48550/arxiv.2103.01745
• Publication Date: 2021-01-01
• Authors: Yunqian Wen; Song Li; Bo Liu; Ming Ding; Rong Xie
• Keywords: Differential Privacy; Identification

Because of the explosive growth face photos as well their widespread dissemination and easy accessibility in social media, security privacy personal identity information becomes an unprecedented challenge. Meanwhile, convenience brought by advanced identity-agnostic computer vision technologies is attractive. Therefore, it important to use images while taking careful consideration protecting people's identities. Given a image, de-identification, also known anonymization, refers generating another image with similar appearance same background, real hidden. Although extensive efforts have been made, existing de-identification techniques are either insufficient photo-reality or incapable well-balancing utility. In this paper, we focus on tackling these challenges improve de-identification. We propose IdentityDP, anonymization framework that combines data-driven deep neural network differential (DP) mechanism. This encompasses three stages: facial representations disentanglement, $\epsilon$-IdentityDP perturbation reconstruction. Our model can effectively obfuscate identity-related faces, preserve significant visual similarity, generate high-quality be used for tasks, such detection, tracking, etc. Different from previous methods, adjust balance utility through budget according pratical demands provide diversity results without pre-annotations. Extensive experiments demonstrate effectiveness generalization ability our proposed framework.