ABSTRACT

Automated facial recognition is increasingly used to identify individuals for various applications. However, face morphing attacks, which blend facial images of two or more people, present a security risk. Morph Attack Detection (MAD) systems have been developed in response to this threat, as Face Recognition Systems (FRS) are vulnerable to such attacks. Because of the limited number of publicly available face morph datasets to investigate, especially to our knowledge, there is no Kurdish morph dataset. In this work, we generated a new face dataset, which we called “KurdFace dataset,” containing morphed face images.  We investigate the susceptibility of biometric systems to morphed face attacks using a morph attack detection model based on traditional machine learning and deep-learning methods to differentiate between genuine and morphed images. In this work, we used Local Binary Patterns (LBP) and uniform Local Binary Patterns (U-LBP) as one block (without blocking), LBP and U-LBP with multi-blocking (dividing cropped frontal face images into different sub-regions). For classification purpose, we utilized the support vector machine (SVM) as a classifier. On the other hand, we used four pre-trained convolutional neural networks (CNN) in a Transfer Learning (TL) mode, to see how they work on morphed images.

Experiments are conducted on Advanced Multimedia Security Lab (AMSL) dataset – which is a publicly available morph dataset – and our newly constructed dataset with two different data separation scenarios, and we considered the balanced and unbalanced issues in our experiments. As a result, the multi-blocking strategy dramatically reduced the classification error rate in all experiments on both datasets. It is worth mentioning that deep features obtained a 0% classification error rate in some cases.