Integrating Diversity into Neural-Network-Based Face Deidentification

Abstract

The extensive use of video surveillance and mass visual media distribution in modern video sharing, social media and cloud services have ignited concerns about the privacy of people identifiable in the scenes. A popular approach for visual privacy protection is face deidentification technique, which aims at concealing the identity of individuals in the video or image while still preserving certain facial attributes after deidentification. Prior research has successfully proposed κ-Same solutions, which provide k-anonymity privacy protection. However, they always cause every $k$ original faces to share the same de-identified faces, making the diversity within the set of the original faces lost. To address this challenge, this paper presents an deidentification procedure, which ensures anonymity by synthesizing surrogate faces for deidentification using generative neural networks (GNNs) while integrates diversity into de-identified faces. We demonstrate the feasibility of the proposed method with extensive experiments which show that our alteration method is indeed effective.