FUSING TEXT, IMAGES, AND USER METADATA: A HYBRID APPROACH TO FAKE NEWS DETECTION ON TWITTER

Abstract

Such extensive spread of misinformation via social media is a significant threat to trustworthiness, the health of the citizens, and healthy democratic institutions. Conventional methods of fake news detection are more or less limited to textual analysis and many of the time the multimodal context abundant on the web is neglected. In this research thesis, a multi-modal fake news detection framework which synergistically incorporates natural language processing and visual processing based on computer vision segments, along with user metadata analytic will be proposed to improve the accuracy of fake news detection on Twitter. The proposed system makes use of a DistilBERT model-based text analysis of the tweet content, ResNet 18-based convolutional neural network based on attached images, and feature-based knowledge of credibility based on account verification status, the number of followers, and account age. The products of these separate elements are combined into final binary classification into REAL or FAKE boundaries with the help of a Random Forest meta-classifier. The whole pipeline is implemented as a Flask web application, which allows in real-time analyzing tweet URLs. It is seen in experimental tests on both synthetic and real world datasets that the system shows strong returns in conditions of different inputs, such as when one or both modality do not work (e.g. a tweet only contains text or only an image). The suggested method provides an below-high-fidelity, socially scalable, and trusted measure of addressing the effects of misinformation between multimodal community media environments.