De-identification Without Losing Faces

Abstract

Training of deep learning models for computer vision requires large image or video datasets from real world. Often, in collecting such datasets, we need to protect the privacy of the people captured in the images or videos, while still preserve the useful attributes such as facial expressions. In this work, we describe a new face de-identification method that can preserve essential facial attributes in the faces while concealing the identities. Our method takes advantage of the recent advances in face attribute transfer models, while maintaining a high visual quality. Instead of changing factors of the original faces or synthesizing faces completely, our method use a trained facial attribute transfer model to map non-identity related facial attributes to the face of donors, who are a small number (usually 2 to 3) of consented subjects. Using the donors' faces ensures that the natural appearance of the synthesized faces, while ensuring the identity of the synthesized faces are changed. On the other hand, the FATM blends the donors' facial attributes to those of the original faces to diversify the appearance of the synthesized faces. Experimental results on several sets of images and videos demonstrate the effectiveness of our face de-ID algorithm.

AnonFACES

“…In this appendix we compare visual results of AnonFACES with recent results from [18] and [7]. As demonstrated in Figure 17, the results in [18] have a blurring effect which reduces the image quality while some identity features (e.g.…”

“…general facial landmark) are still similar. AnonFACES clearly outperforms [18] in terms of naturalness and has a noticeably better anonymizing effect. In Figure 18 the results from [7] are shown to have minimal difference in terms of identity features such as general facial landmark features, eyes, and mouth area.…”

“…This is important as it allows, for example, car manufacturers to effectively train their autonomous vehicles while complying with privacy laws. [18]: first row: input images, second row: results from [18], third row: AnonFACES results using =2 Figure 18: Comparison with Gafni et al [7]: first row: input images, second row: results from [7], third row: AnonFACES results using =2…”

“…In both cases, our results still preserve non-identity information such as emotion, age, gender, and skin color. Note, however, that [18] and [7] only replace some (identifying) facial features, presumably to work with videos, whereas AnonFACES replaces the entire image. Restricting the area to within the face would allow for a more direct comparison but is outside of the scope of this paper.…”

Learning Semi-Supervised Anonymized Representations by Mutual Information

“…De-identification is similar to anonymization but consists in the process of preventing someone personal ID from being revealed. The main difference between this concept and data anonymization is that some identifying information can be preserved in order to be relinked only by a trusted party or by the original data operator, whereas in the case of anonymization, no re-identification should be possible (e.g., [6,7]).…”