Hiding Among Your Neighbors: Face Image Privacy Protection with Differential Private k-anonymity

Abstract

The development of modern social media allows millions of private photos to be uploaded and shared, which provides a wide range of image acquisition but extremely threatens personal image privacy. Face de-identification is treated as an important privacy protection tool in multimedia data processing by modifying image identity information. Although there exist many traditional methods widely used to hide sensitive private information, they all fail to balance the trade-off between privacy and utility in qualitative and quantitative manners and cannot generate de-identified results with satisfactory visual perception. In this paper, we propose a novel face image privacy protection method with differential private k-anonymity, which can not only generate de-identified results with good image quality but also control the balance between privacy protection and image utility according to different application scenarios. The framework consists of the following three steps: facial attributes prediction, privacy-preserving attributes obfuscation, and naturally realistic de-identificated image generation. Our extensive experiments demonstrate the stability and effectiveness of the proposed model.