Hukkelås Håkon ; Mester Rudolf ; Lindseth Frank

Abstract

We propose a novel architecture which is able to automatically anonymizefaces in images while retaining the original data distribution. We ensure totalanonymization of all faces in an image by generating images exclusively onprivacy-safe information. Our model is based on a conditional generativeadversarial network, generating images considering the original pose and imagebackground. The conditional information enables us to generate highly realisticfaces with a seamless transition between the generated face and the existingbackground. Furthermore, we introduce a diverse dataset of human faces,including unconventional poses, occluded faces, and a vast variability inbackgrounds. Finally, we present experimental results reflecting the capabilityof our model to anonymize images while preserving the data distribution, makingthe data suitable for further training of deep learning models. As far as weknow, no other solution has been proposed that guarantees the anonymization offaces while generating realistic images.