In this thesis, we present two novel approaches to de-identify visual data. They leverage Generative Adversarial Networks and Diffusion Models, two recently developed generative deep learning techniques, to replace real faces with synthetically generated surrogates. The advantage of this approach over traditional anonymization with pixelization or blurring is that it can retain data utility for downstream tasks that require processing the human face. Our first approach, DetailedPrivacy, can preserve expression, pose and gaze of individual faces but can only be applied to images and videos containing relatively large faces with little occlusion. Our second approach, on the other hand, StablePrivacy, can be applied to more complex scenes and alters faces more drastically. It achieves state-of-the-art protection against identification by deep learning-based face recognition methods. Moreover, it retains the utility necessary for training deep learning-based face detection models on anonymized data better than all other approaches we evaluated.     «

In this thesis, we present two novel approaches to de-identify visual data. They leverage Generative Adversarial Networks and Diffusion Models, two recently developed generative deep learning techniques, to replace real faces with synthetically generated surrogates. The advantage of this approach over traditional anonymization with pixelization or blurring is that it can retain data utility for downstream tasks that require processing the human face. Our first approach, DetailedPrivacy, can pres...     »