On the detection of GAN-based face morphs using established morph detectors

Abstract

Face recognition systems (FRS) have been found to be highly vulnerable to face morphing attacks. Due to this severe security risk, morph detection systems do not only need to be robust against classical landmark-based face morphing approach (LMA), but also future attacks such as neural network based morph generation techniques. The focus of this paper lies on an experimental evaluation of the morph detection capabilities of various state-of-the-art morph detectors with respect to a recently presented novel face morphing approach, MorGAN, which is based on Generative Adversarial Networks (GANs). In this work, existing detection algorithms are confronted with different attack scenarios: known and unknown attacks comprising different morph types (LMA and MorGAN). The detectors’ performance results are highly dependent on the features used by the detection algorithms. In addition, the image quality of the morphed face images produced with the MorGAN approach is assessed using well-established no-reference image quality metrics and compared to LMA morphs. The results indicate that the image quality of MorGAN morphs is more similar to bona fide images compared to classical LMA morphs.