The Eyes: A Source of Information for Detecting Deepfakes

Currently, the phenomenon of deepfakes is becoming increasingly significant, as they enable the creation of extremely realistic images capable of deceiving anyone thanks to deep learning tools based on generative adversarial networks (GANs). These images are used as profile pictures on social media with the intent to sow discord and perpetrate scams on a global scale. In this study, we demonstrate that these images can be identified through various imperfections present in the synthesized eyes, such as the irregular shape of the pupil and the difference between the corneal reflections of the two eyes. These defects result from the absence of physical and physiological constraints in most GAN models. We develop a two-level architecture capable of detecting these fake images. This approach begins with an automatic segmentation method for the pupils to verify their shape, as real image pupils naturally have a regular shape, typically round. Next, for all images where the pupils are not regular, the entire image is analyzed to verify the reflections. This step involves passing the facial image through an architecture that extracts and compares the specular reflections of the corneas of the two eyes, assuming that the eyes of real people observing a light source should reflect the same thing. Our experiments with a large dataset of real images from the Flickr-FacesHQ and CelebA datasets, as well as fake images from StyleGAN2 and ProGAN, show the effectiveness of our method. Our experimental results on the Flickr-Faces-HQ (FFHQ) dataset and images generated by StyleGAN2 demonstrated that our algorithm achieved a remarkable detection accuracy of 0.968 and a sensitivity of 0.911. Additionally, the method had a specificity of 0.907 and a precision of 0.90 for this same dataset. And our experimental results on the CelebA dataset and images generated by ProGAN also demonstrated that our algorithm achieved a detection accuracy of 0.870 and a sensitivity of 0.901. Moreover, the method had a specificity of 0.807 and a precision of 0.88 for this same dataset. Our approach maintains good stability of physiological properties during deep learning, making it as robust as some single-class deepfake detection methods. The results of the tests on the selected datasets demonstrate higher accuracy compared to other methods.