FACE MORPHING ATTACK DETECTION IN THE PRESENCE OF POSTPROCESSED IMAGE SOURCES USING AVERAGING DIMENSIONALITY REDUCTION AND FEATURE-LEVEL FUSION

Abstract

Face Morphing Attack Detection (MAD) has recently received a lot of attention because criminals have started to merge two or more subject facial images using publicly and widely obtainable digital manipulation techniques to develop a new facial image that can be interpreted as an accurate image of any of the individual images that make it up. Some of these tools generate high quality morphed images that pose a significant challenge to existing Face Recognition Systems (FRS). FRS has been shown to be vulnerable to multiform morphing attacks in the literature.Several forms of research on the detection of this morph attack have been carried out on the basis of this vulnerability using several techniques. Despite the high levels of MAD recorded in the literature, no suitable solution for handling post-processed pictures, such as those updated after morphing with a sharpening operation that significantly reduces visible artefacts in morphed photos, has yet to be discovered. In this work, before image post-processing and after image postprocessing, an approach is proposed for MAD based on averaging dimensionality reduction and feature-level fusion and classification using Support Vector Machine (SVM). The outcome of SVM training with fused feature vectors increased the accuracy of the classification from 94% to 97.1%, thus enhancing overall performance.