Fusion of Deep Features for Differential Face Morphing Attack Detection at Automatic Border Control Gates

Abstract

Face recognition systems (FRS) are showing increasing accuracy in an uncontrolled real world, leading to their usage in automated border control (ABC) gates. However, both automatic and manual FRS are prone to Face Morphing Attacks (FMA), which can be generated by linearly blending face images from two contributory data subjects. Differential Morphing Attack Detection (D-MAD), which compares the face image in an electronic Machine Readable Travel Document (eMRTD) with a trusted live capture from an ABC gate, is thus a significant problem to address. This paper presents Robust D-MAD (RD-MAD), which performs pair-wise pose normalization on the bona fide and probe face images based on global affine alignment. The proposed D-MAD technique is based on the comparison score level fusion of deep features extracted using off-the-shelves pre-trained deep networks such as AlexNet and ResNet. The deep features are extracted corresponding to both enroled and probe face images independently from AlexNet and ResNet. Then the signed difference of the features is computed between enroled and probe face images independently on both AlexNet and ResNet. Two linear SVMs are trained on the signed difference features corresponding to deep networks whose comparison scores are fused using the weighted sum rule to make the final decision. Extensive experiments are performed on a challenging dataset having lighting, face pose, expression, illumination and print-scan variations. Obtained results outperform the state-of-the-art (SOTA) as we achieve an EER=2.1% compared to the SOTA with an EER=4.6%.