Detecting Deepfakes Using GAN Manipulation Defects in Human Eyes

The Deepfake phenomenon is very important nowadays because there are possibilities to create very real images that can fool anyone, thanks to deep learning tools based on generative adversarial networks (GAN). These images are used as profile images on social media, aimed here at creating discord and scams internationally. In this work, we show that these images can be detected by a multitude of imperfections present in the synthetized eyes such as the irregular shape of the pupil and the difference between the corneal reflections of the two eyes. These imperfections are caused by the absence of physical/physiological constraints in most GAN models. We are developing a two tier architecture able of detecting these deepfake images. It starts with an automatic segmentation method of the eye pupil to check the shape. Then, for pupils of non-standard shape, the whole image is taken, transformed into gray level and then passed into an architecture that extracts and compares the corneal specular reflections of two eyes. Experimenting with a large set of real image data from the Flickr-Faces-HQ dataset and fake styleGAN2 images demonstrates the effectiveness of our method. Our method has good stability for physiological properties during deep learning; therefore, it is robust as some of the single-class deepfake detection methods. The results of the experiments on the selected datasets demonstrate greater precision compared to other methods.