Photorealistic Face De-Identification By Aggregating Donors' Face Components

With the adoption of pervasive surveillance systems and the development of efficient automatic face matchers, the question of preserving privacy becomes paramount. In this context, automated face de-identification is revived. Typical solutions based on eyes masking or pixelization, while commonly used in news broadcasts, produce very unnatural images. More sophisticated solutions were sparingly introduced in the literature, but they fail to account for fundamental constraints such as the visual likeliness of de-identified images. In contrast, we identify essential principles and build upon efficient techniques to derive an automated face de-identification solution meeting our predefined criteria. More specifically, our approach relies on a set of face donors from which it can borrow various face components (eyes, chin, etc.). Faces are then de-identified by substituting their own face components with the donors' ones, in such a way that an automatic face matcher is fooled while the appearance of the generated faces are as close as possible to original faces. Experiments on several datasets validate the approach and show its ability both in terms of privacy preservation and visual quality.