Creating a defense against face de-identification

Face recognition systems are used to identify a person’s face by detecting patterns in their facial features. It is possible to throw off the FR system by deliberately obfuscating the facial patterns the FR system looks for, while at the same time keeping the image unchanged to the human eye. Due to the large range of potential adversarial attack methods, the goal of this research is to improve robustness of FR systems by training the model using obfuscated images as well as regular images, instead of just regular images.