Detection of Visual Deepfakes using Deep Convolutional Networks

Abstract

With social media being a significant part of everyday life, the possibilities of misinformation spreading in textual, visual or audio form are now significantly in comparison to past. With the onset of widely available generative AI models, the need for effective classification methods for generated or altered content grew even larger. This article focuses on the problem of Deepfake detection, particularly in a domain of artificially generated depictions of human faces. For the detection, we have selected a variety of CNN (Convolutional Neural Networks) based architectures which have recently proven their capabilities in image classification tasks. We have selected five models and performed testing on a two-class dataset which contained Deepfakes created with state-of- the-art StyleGAN. The achieved results of selected models were comparable, each model attaining sufficient classification capability.