Visual Feature Space Analyses of Face Morphing Detectors

Abstract

Face morphing attacks are a serious threat to biometric face recognition systems. If a morphed face image containing biometric characteristics of at least two different subjects is embedded into a personal document, this document can be used for verification by different subjects. Due to the fast dissemination of automated face recognition systems, especially for border control at airports, detection of morphed face images in documents is gaining relevance. Recently proposed morphing detection approaches demonstrate low error rates with datasets used for development and evaluation, but often fail for new datasets. The detectors are often trained and evaluated on images in better quality than images in electronic Machine-Readable Travel Documents (eMRTD). Since these morphing detectors are clearly suitable for fraud detection before the image is compressed to fit into a eMRTD, we study their performance and potential on images compliant with eMRTDs. A visual analysis of feature spaces provide us with a better understanding of a presumable generalization ability of morphing detectors. The questions we endeavour to answer are how the distributions of high resolution face images and face images in eMRTDs differ in a detector’s feature space and whether detecting morphed face images in eMRTDs should be considered as a different task than detecting them before they are compressed to fit into an eMRTD.