Towards Robust Evaluation of Face Morphing Detection

Automated face recognition is increasingly used as a reliable means to establish the identity of persons for various purposes, ranging from automated passport checks at the border to transferring money and unlocking mobile phones. Face morphing is a technique to blend facial images of two or more subjects such that the result resembles both subjects. Face morphing attacks pose a serious risk for any face recognition system. Without automated morphing detection, state of the art face recognition systems are extremely vulnerable to morphing attacks. Morphing detection methods published in literature often only work for a few types of morphs or on a single dataset with morphed photographs. We create face morphing databases with varying characteristics and how for a LBP/SVM based morphing detection method that performs on par with the state of the art (around 2% EER), the performance collapses with an EER as high as if it is tested across databases with different characteristics. In addition we show that simple image manipulations like adding noise or rescaling can be used to obscure morphing artifacts and deteriorate the morphing detection performance.