Context-Aware Multimodal Fake News Detection using Localized DistilBERT

Using deep learning frameworks to detect fake news is a sophisticated step toward automating the verification of media content, helping to identify and avert misinformation with speed and efficiency. In this work, we developed an end-to-end localized workflow for the binarization of news articles with a specially designed MultimodalBERT model. Our approach classifies news samples as "fake" and "real" using various modalities, such as the text of the article, the article’s source, user activity metrics, and other documents. The dataset has rich contextual diversity because it is comprised of records from FakeNewsNet, LIAR, and FEVER. All tasks completed in our system are done offline, locally with Python scripts in PyTorch, and without using any external cloud computational resources. In accordance with explainable AI principles, our approach allows users to understand model confidence and score distribution alongside the model’s prediction. The model displays outstanding performance with an F1 score of 99.3%.