CHIEFS: Corneal-Specular Highlights Imaging for Enhancing Fake-Face Spotter

This paper presents a novel Machine Learning (ML)-based DeepFake detection technology named CHIEFS (Corneal-Specular Highlights Imaging for Enhancing Fake-Face Spotter). We focus on the most reflective area of a human face, the eyes, upon the hypothesis that the existing DeepFake creation methods fail to coordinate their counterfeits with the reflective components. In addition to the traditional checking of the reflection shape similarity (RSS), we detect various corneal-specular highlights features, such as color components and textures, to find corneal-specular highlights consistency (CHC). Furthermore, we inspect the ensemble of the highlights with the surrounding environmental factors (SEF), including the light settings, directions, and strength. We designed and built them as modular features and have conducted extensive experiments with different combinations of the components using various input parameters and Deep Neural Network (DNN) architectures on Generative Adversarial Network (GAN)-based DeepFake datasets. The empirical results show that CHIEFS with three modules improves the accuracy from 86.05% (with the RSS alone) to 99.00% with the ResNet-50-V2 architecture.