ABSTRACT
Collecting and storing images has become necessary for many applications, especially for computer vision development (e.g. in the development of autonomous vehicles). However, freely collecting images violates privacy regulations in several regions, unless the data are anonymised. Traditional image anonymisation (e.g. blurring) degrades the original data, making the images unusable for many applications. Recent developments in deep generative models have enabled a new type of anonymisation: realistic anonymisation. This new technology replaces privacy-sensitive information with artificially synthesised, realistic content. These methods preserve individuals’ privacy and generate visually pleasing images almost indistinguishable from the original data.
