Morphing detection based on regional analysis of local frequency content

Face recognition software is known to be vulnerable to a presentation attack in the form of face morphing. Face morphing detection is an active field of research. Creat- ing strong face morphing detection algorithms will make face recognition software more robust. This paper inves- tigates how regional analysis of the frequency spectrum of face images can be used to detect morphs in both a dif- ferential and non-differential setting. Three methods are explained and assessed for their performance. The first method utilizes the Kullback Leibler Divergence. The sec- ond is a Support Vector Machine (SVM). The third a Deep Feed Forward Neural Network (DFF). The latter two are trained on the frequency spectrum. The Kullback Leibler Divergence proved to be not discriminate enough to clas- sify morphs. Both the SVM and DFF were able to detect morphs with an accuracy of around 80%.