MKEE: Multimodal and Knowledge-Enhanced Ensemble for Fake News Detection

Abstract

The rapid spread of misinformation on digital platforms has created an urgent need for reliable and scalable fake news detection systems. Traditional text-only methods are insufficient against modern misinformation campaigns that cleverly mix misleading language, altered images and false factual context. This paper presents the Multimodal Knowledge-Enhanced Ensemble (MKEE) framework, which combines text, visuals and knowledge-based signals within a single classification process. The system has three cooperative stages: it extracts cross-modal features using RoBERTa and CLIP encoders, grounds facts through a Heterogeneous Graph Transformer (HGT) that queries Wikidata and DBpedia, and performs lightweight ensemble classification with XGBoost. Our proposed framework significantly boosts resilience against complex misinformation while keeping computational costs within deployment limits. The experimental results show that the proposed work outperforms the existing model in terms of accuracy, precision and F1-Score. A key design choice in this architecture is the separation of feature extraction from classification, which helps avoid the scalability issues found in fully end-to-end deep learning systems.