Privacy preservation through facial de-identification with simultaneous emotion preservation

Abstract

Due to the availability of low-cost internet and other data transmission media, a high volume of multimedia data get shared very quickly. Often, the identity of individuals gets revealed through images or videos without their consent, which affects their privacy. Since face is the only biometric feature that reveals the most identifiable characteristics of a person in an image or a video frame, the need for the development of an effective face de-identification algorithm for privacy preservation cannot be over-emphasized. Existing solutions to face de-identification are either non-formal or are unable to obfuscate identifiable features completely. In this paper, we propose an automated face de-identification algorithm that takes as input a facial image and generates a new face that preserves the emotion and non-biometric facial attributes of a target face. We consider a proxy set of a large collection of artificial faces generated by StyleGAN and select the most appropriate face from the proxy set that has a facial expression and pose similar to that of the target face. The faces in the proxy set are artificially generated, and hence the face selected from this set is completely anonymous. To retain the non-biometric attributes of the target face, we employ a generative adversarial network (GAN) with a suitable loss function that fuses the non-biometric attributes of the target face with the face selected from the proxy set to obtain the final de-identified face. Experimental results emphasize the superiority of our approach over state-of-the-art face de-identification methods. © 2020, Springer-Verlag London Ltd., part of Springer Nature.