Morphing Attack Detection is a current field of research. Morphing is the targeted combination of several images, for example of human faces. These morphed images can be used during the passport application process. All
subjects contained in the morph can be recognised by existing face recognition systems and as a result (automatic) border controls can be circumvented. A number of studies have shown that existing face recognition systems and
human experts can be deceived by morphed images. Based on this finding, various approaches for morphing as well as for the automatic recognition of morphing attacks have been published. Automated morphing detection is
still a young branch of research with many recent publications.
During the previous internship we were given features extracted from various images using different feature extractors: ArcFace, FaceNet, Eyedea,
LM-Wing, LM-Dlib, LBP13, LBP43, BSIF13 and BSIF43. This thesis will focus on developing a fusion approach to Morphing Attack Detection using scores generated based on different classifiers with optimised hyperparameters. Four different algorithms were used for classification, namely: Support Vector Machines, Random Forest, AdaBoost and GradientBoosting. The hyperparameters were optimised in three different ways: using grid-search, an evolutionary approach, and bayesian optimisation. We used different fusion techniques and compared their results. Our focus in this work was score-level fusion and sum-rule – an equally weighted sum-rule and two non-equally weighted sum-rules were used for fusion. One of the non-equally weighted approaches finds its weights using grid-search and the other using random forest.
We noticed that not using equal weights achieves better results and even though grid-search weights might lead to better results than random forest weights, grid-search is more time-consuming. However, both random forest and grid-search weights can significantly improve the Detection Equal Error Rate.