Résumé
With the growing storage and diffusion of multimedia data across digital networks, protecting visual content is a subject of interest in research, especially via the means of image obscuration methods. Although several of these techniques have been explored to provide basic to advanced protection against re-identification by humans or automated recognition systems, few achieve full key-based reversibility while introducing minimal distortion to the resulting image. In this paper, we propose a novel image content obscuration method that leverages variational autoencoders to address these limitations of obscuration methods. Our approach transforms high-dimensional images belonging to a source class into lower-dimensional representations (latent vectors), and then applies three distinct transformations to the latent representation in order to match it to the target class. These transformations are designed to be both visually imperceptible and reversible using a secret key, enabling the original content to be accurately reconstructed. We evaluate our method through qualitative and classification-based experiments regarding the obscuration and defense against re-identification, and compare to previous image obscuration methods.