On Improving Face Generation for Privacy Preservation

Replacing faces in image and video content with generated ones (e.g., using generative adversarial networks, GANs) has gained attention recently, as it enables resolving privacy issues in image and video data used for visualization purposes or training data in multimedia analysis and retrieval systems. Privacy issues should be addressed when visual content enters the system, as identifying and removing content later (which may be necessary due to the shifts in legislation and users' increased awareness) is a tedious and costly task. This paper proposes two improvements of face generation: First, we propose the use of portrait segmentation on the training data of the GAN, in order to generate images that are not only cropped to the face region, which may cause artifacts during insertion. Second, we add a face detection term to the loss function, in order to better guide the training process. The results show that these modifications enable creating uncropped face images achieving the same or better performance than for closely cropped images. We use the detectability of the generated faces as an evaluation metric, discuss the limitations of such a metric and propose enhancements for better comparability. We also demonstrate that the aim of anonymization is achieved by running face recognition on the modified images from the LFW data set.