HMFND-NET: A HYBRID MULTIMODAL FAKE NEWS DETECTION NETWORK USING XLNET, BILSTM AND ATTENTION-BASED FEATURE FUSION

Abstract

Fake news dissemination through digital platforms has emerged as a major societal challenge, adversely influencing public opinion, political processes, financial markets, and healthcare decisions. Although recent advancements in deep learning and transformer-based architectures have significantly improved fake news detection performance, existing approaches continue to suffer from critical limitations related to contextual understanding, sequential dependency modeling, and intelligent feature selection. Transformer models provide rich contextual representations but may inadequately capture sequential dependencies, whereas recurrent neural networks effectively model temporal relationships but lack advanced contextual learning capabilities. To address these complementary limitations, this paper proposes a novel Hybrid Multimodal Fake News Detection Network (HMFND-Net) that integrates XLNet contextual embeddings, Bidirectional Long Short-Term Memory (BiLSTM) sequence learning, and Attention-based feature fusion within a unified end-to-end trainable architecture. The proposed framework first generates contextual embeddings using XLNet and subsequently applies BiLSTM to capture bidirectional sequential dependencies. An attention mechanism is then employed to identify and emphasize the most informative features while suppressing irrelevant information. A multi-stream feature fusion layer combines contextual, sequential, and attention-weighted representations for robust fake news classification. HMFND-Net was comprehensively evaluated using three benchmark datasets: LIAR, ISOT, and WELFake. Comparative experiments were conducted against nine baseline models spanning Machine Learning, Deep Learning, and Transformer-based paradigms. Experimental results demonstrate that HMFND-Net consistently outperforms all existing approaches, achieving 99.1% accuracy, 99.0% precision, 98.9% recall, 98.95% F1-score, 0.99 MCC, and 0.995 ROC-AUC. Ablation studies confirm the contribution of each architectural component, and statistical significance testing (p < 0.05) validates the superiority of the proposed framework. Cross-dataset evaluation demonstrates strong generalization capability (95.9%–97.2%), and robustness analysis confirms performance exceeding 96.9% under diverse noisy conditions. The integration of contextual learning, sequential modeling, and attention-guided feature selection significantly enhances fake news detection performance and establishes HMFND-Net as an effective framework for intelligent misinformation identification.