Detecting soft-biometric privacy enhancements on face images

Abstract

In scientific literature, there is a growing need for methods to ensure the privacy in digital images. In the field of face analytics, researchers have proposed privacy-preserving techniques which transform face images in such a way that the automatic extraction ofsoft-biometrics (e.g., gender) is prevented, while the visual appearance gets minimally degraded. We present a novel technique todetect whether or not an image was manipulated with such privacy-preserving techniques. Our detector exploits the fact that the soft–biometric classifier gives different results for privacy-enhanced images and their reconstructed versions. In our experiments, we have demonstrated that this difference can be exploited to detect whether an image was privacy-enhanced. The advantage of our method is that the used privacy-enhancing technique does not need to be known in advance (black-box scenario). Our approach is evaluated considering four privacy-enhancing techniques for soft-biometrics on three versatile face datasets. The results show that our approach has a number of advantages over competing techniques and that it detects privacy-enhancement with high accuracy.