Progress in deep learning-based computer vision has been significantly accelerated by the surge in available datasets for training and testing. However, the failure to meetethical and regulatory standards in datasets containing privacy-sensitive content such as facial images has caused public concern and even led to the withdrawal of datasets. While traditional anonymization strategies, such as pixelization, offer a seemingly straight forward solution, they lack the ability to maintain the necessary facial details crucial for applications like training face detection models. To reconcile the need for high quality data with stringent privacy standards, we explore an innovative method for de identification that employs Stable Diffusion using synthetically generated faces as image prompts alongside a noisy version of the original face to guide anonymization, which we term StablePrivacy. Our experiments demonstrate the capability to preserve detailed features for training high-quality face detection models while offering state-of-the-art privacy protection.
Leibl et al. (Thu,) studied this question.