Roman Kessler successfully defended his Master’s Thesis

Roman Kessler successfully defended his Master’s Thesis on ‘Analysis of face embeddings to facilitate image pre-selection for face morphing’

Face Morphing Attacks pose a novel threat to the security of identification documents. The fusion of the face images of two or more – similarly look- ing – individuals during the application process for a new travel document (i.e., passport) or identity card enables both individuals to travel with the same document. In order to develop algorithms to detect morphing attacks, large data sets of morphed face images are needed, for which in turn many similarly looking individuals need to be paired.

The study at hand uses face embeddings of openly accessible face recog- nition models to describe similarity between individuals. It aims at finding appropriate face recognition models, metrics to quantify similarity, morph- ing algorithms to fuse facial images of paired individuals, and soft biometric characteristics to analyze the attack potential of face morphs.

Results demonstrate, that image pre-selection based on Cosine or Euclidean distances between face embeddings highly improves the attack potential of morphs. Especially ArcFace and MagFace provide valuable face embeddings to quantify similarity for pre-selection. Both open source, as well as Commer- cial Off-The-Shelf Face Recognition Systems get fooled by morphed faces. Landmark-based, closed source morphing algorithms pose high risk for any of the tested Face Recognition Systems. On the other hand, MagFace embed- dings further emerge as valuable means to detect morphed face images. Soft biometrics characteristics however were only partially relevant to predict morph success, if morphing has been conducted within similar age, gender, and race groups.

The results emphasize that face embeddings are valuable instruments on both sides of the morphing attack, image pre-selection for face morphing and detection of morphed faces.

Roman Kessler successfully defended his Master’s Thesis