Our paper “On the Importance of Conditioning for Privacy-Preserving Data Augmentation” by Julian Lorenz, Katja Ludwig, Valentin Haug, and Rainer Lienhart has been accepted to the Workshop on Safe and Trustworthy Multimodal AI Systems (SafeMM-AI) at the International Conference on Computer Vision 2025. Additionally, our paper has been selected for an oral presentation at the conference by the reviewers.

In this publication, we examine a method for anonymizing faces and identities. Contrary to initial claims, we were able to bypass the anonymization and accurately identify the original person with a success rate of 69%. We attribute this vulnerability to the fact that the anonymization method attempts to maintain the basic structure of the image. Although the results appear very convincing to the human eye, a neural network can successfully bypass the anonymization. We conclude that any structure related to the original image must be removed to ensure a successful anonymization.