Fake News Detection via Embedded Visual Content

Abstract

In recent years, given the prevalence of misinformation and fake news circulating online, in particular on Social Media, the demand for computational tools assessing the veracity of news increased. Using a Design Science approach, this thesis focused on a content-based fake news detection solution by identifying images alteration techniques that were at times used in the creation of fake news articles. In this thesis we proposed two artifacts, FaRe-PS and FaRe-GAN for the detection of two common image manipulation techniques: image splicing and image generation using General Adversarial Networks. Through the use of ensemble modeling technique and the current state-of-the-art image classification algorithm, convolutional neural network, the proposed neural networks achieved promising results. The ensemble model for the detection of image splicing, FaRe-PS was capable of detecting altered images with an accuracy of 65,05% and an AUROC score of 0,695. While, the detection of machine generated images with the devised model, FaRe-GAN, on the test images achieved an accuracy of 96,81% and an AUROC score of 0,986. To benchmark the capability of the two artifacts on the task of fake facial image detection, the judgement of the human perception was used, by means of survey evaluation. In comparison to FaRe-PS, evaluators managed to detect spliced images with a comparable accuracy of 65%. While, the developed FaRe-GAN model has far outperformed evaluators, who were deceived by generated images 43% of the time. The results demonstrated that there is a need for automated fake image detection tool and that the artifacts to a large extent effectively detect the use of two image alteration techniques used in the production of fake news articles.