Exposing GAN-generated Faces Using Inconsistent Corneal Specular Highlights

Sophisticated generative adversary network (GAN) models are now able to\nsynthesize highly realistic human faces that are difficult to discern from real\nones visually. In this work, we show that GAN synthesized faces can be exposed\nwith the inconsistent corneal specular highlights between two eyes. The\ninconsistency is caused by the lack of physical/physiological constraints in\nthe GAN models. We show that such artifacts exist widely in high-quality GAN\nsynthesized faces and further describe an automatic method to extract and\ncompare corneal specular highlights from two eyes. Qualitative and quantitative\nevaluations of our method suggest its simplicity and effectiveness in\ndistinguishing GAN synthesized faces.\n