A Hybrid Model for Identity Obfuscation by Face Replacement

Abstract

As more and more personal photos are shared and tagged in social media, avoiding privacy risks such as unintended recognition, becomes increasingly challenging. We propose a new hybrid approach to obfuscate identities in photos by head replacement. Our approach combines state of the art parametric face synthesis with latest advances in Generative Adversarial Networks (GAN) for data-driven image synthesis. On the one hand, the parametric part of our method gives us control over the facial parameters and allows for explicit manipulation of the identity. On the other hand, the data-driven aspects allow for adding fine details and overall realism as well as seamless blending into the scene context. In our experiments we show highly realistic output of our system that improves over the previous state of the art in obfuscation rate while preserving a higher similarity to the original image content.

Live Face De-Identification in Video

“…In addition, the work of [49] presents results on low-resolution black and white images only, with no pose or gender variation. Figure 5 compares with the recent work of [43,44]. Our method is able to distance the identity in a more subtle way, while introducing less artifact.…”

“…We provide an extensive comparison with the work of [44]. In the paper we only included one of [44] generated outputs: the sample shown was the first output that gains <50% of recognition rates by an automatic face recognition algorithm, according to [44]. The work of [44] provides several models for different levels of de-identification.…”

“…However, the experiments were restricted to a homogeneous dataset, with no apparent expression preservation within the results. In the GAN-based methods of [43,44], face de-identification is employed for the related task of person obfuscation. The work of [43] conditions the output image based on both a blurred version of the input and the extracted facial pose information.…”

“…(a) (b) (c) Figure 5. (a) Input images from [43,44], (b) our results, (c) those of [43] (row 1) and [44] (rows 2-3).…”

“…As far as we know, this is the first time that a representation from an existing classifier network is used to augment an autoencoder, which enables the feed-forward treatment of new persons, unseen during training. In addition, this is the first work to introduce a new Newton, '05 [32] Gross, '08 [10] Samarzija, '14 [41] Jourabloo, '15 [16] Meden, '17 [31] Wu, '18 [49] Sun'18 [43,44] Our Preserves We compare to all methods that provide reasonable quality images in their manuscript, under conditions that are favorable to previous work (we crop the input images from the pdf files, except for the images received from the authors of [43,44]). † The face is swapped with an average of a few dataset faces.…”