Classification of Real and Ai-Generated Fake Faces Using Deep Learning

Abstract

Generative adversarial network has gained popularity mainly due to its ability to create realistic fake human faces. The remarkable detail with which such images have been created in the past few years has exceeded the ability of humans to differentiate between these fake images and real images. Such images have been known to be capable of deceiving the face recognition systems with certain success as well. Forensic systems being developed nowadays take into account adversarial attacks in order to create a more comprehensive detection approach. Different types of GANs such as StackGAN, StyleGAN, and conditional GANs use different architectures to produce images. Since the underlying technology is different from one another it is difficult for any single detection algorithm trained on one kind of GAN to detect fake images generated from some other kind of GAN. This research is intended to address this apparent lack of a generalized solution for the detection of GAN generated images. In the proposed study, we use images from three different datasets; FFHQ, StyleGAN faces and PGAN faces. A siamese network with triplet loss is proposed that takes three images as inputs: anchor, positive and negative, and makes a prediction based on the similarity score of their embeddings. Extensive experiments have been conducted to analyze the effectiveness of the proposed approach. The results and evaluation show that the siamese triplet loss network performs significantly better than the contemporary approaches with accuracy exceeding 90% for a network trained and tested on images from different datasets