Building robust morphing attacks for face recognition systems

In this paper, a method to build a robust morphing attack to a face verification system will be presented. The proposed method has been developed to investigate the robustness and the impact of morphing attacks to face recognition systems. In this kind of attack, an impostor accesses a face recognition system (FRS) which compares its real-time image with a stored morphed image, built with the impostor and a legitimate user. The attack succeeds when the FRS accepts the impostor and accesses the system. The current approach offers a method to build a robust attack to the FRS, in the sense that the morphed image will be closest to the decision threshold. Morphing attacks are usually evaluated only with images in which both subjects contributed in the same way to the morphing images. The image database considered, FRAV Database, was made up of 200 images. Likewise, two stages were carried out. The first stage was designed to build a baseline reference: an FRS system (trained only with legitimate users) was tested with morphed images. One contribution of this paper is that this test, which usually only considers a 50% fusion between two images has been enriched and some fusion contributions has been considered. Tests have been conducted with 20%, 40%, 50%, 60% and 80% contribution of each image to the morphed image. The comparison of the equal error rate (EER) achieved will show which contribution defines the best plausible attack. Notice that the attack that achieves the best rates with minimum disturbance of the images. The second stage consisted of the reinforcement of the FRS, training it with the contribution set defined in the previous stage. The outcomes obtained achieved improved by 3% of the EER scores.