Deep Residual Learning based Discriminator for Identifying Deepfakes with Cut-Out Regularization

Abstract

The recent development of Generative Adversarial Networks (GANs) have greatly eased the generation of deepfake images which are indistinguishable from real images. As a downside of such advancement, it is now easy to impersonate a person leading to identity theft and other malicious outcomes. In such a scenario it becomes imperative to have a robust algorithm in place which can segregate real images from the fake ones. In this study, we suggest a residual connection based convolutional neural network (CNN) architecture for detecting deepfake images and compare the results with the existing transfer learning algorithms for identifying the deepfakes. The data set used in this study is the combination of the Flickr-Faces-HQ (FFHQ) data set (Nvidia) and the deepfakes generated by the Style GAN, which is proposed by Nvidia. The data set consisting of 1,20,000 images is used for training and validating the network, while a separate set of 20,000 real world images are used for testing the performance of the model. In this current work, we test the robustness of three different algorithms - Inception Resnet V2, VGGFace2, and our customized Residual CNN with and without cut-out regularization in identifying real images. The residual architecture-based implementation in combination with cut-out architecture produces the lowest false positives rate at 0.0043% while the Inception Resnet V2 in combination with cut - regularization produces the best accuracy at 99.05%.