Multi-modal Fake News Detection

Abstract

False information poses a significant risk to society by altering public perception and eroding trust in reliable sources. With the rise of misinformation in various text and image forms, the demand for advanced multi-modal detection systems is growing. This study introduces a multi-modal framework for detecting fake news that integrates NLP, ML, and GNNs, tested on two datasets: a multi-class text-only dataset and a binary text+image dataset. Text embeddings (TF-IDF, GloVe, FastText, BERT) are combined with classifiers (Logistic Regression, Decision Trees, XGBoost, Random Forests) and graph models (GCN, GraphSAGE) to understand structural relationships. An adapted BERT model additionally integrates numerical, categorical, and textual features for enhanced classification. In the binary context, a visual module based on YOLOv11 identifies fake images, and late fusion integrates the top text and vision models into a single classifier. Findings indicate that the suggested framework surpasses unimodal benchmarks: the adjusted BERT secures the highest text accuracy, GraphSAGE excels with graphs, and YOLOv11 fusion enhances detection for visually backed news, underscoring the efficiency of multi-modal fake news detection in practical applications.