XFND: Explainable Fake News Detection using a Hybrid DistillBERT and BiLSTM

The spread of misinformation on social media platforms poses a significant challenge because many people rely on online platforms for updated information and news. Deep learning models, particularly transformer-based black-box models, have demonstrated high accuracy in fake news detection tasks. The main focus of existing studies was specifically on classification performance while neglecting interpretability, which makes these models less transparent and difficult to trust. To address these challenges, a hybrid DistilBERT + BiLSTM model is proposed for fake news detection, using the Fake and Real News dataset. For extracting contextual embeddings, we employ DistilBERT and process them with bidirectional LSTM for sequence modeling. The model achieves the accuracy of 98% which is higher than the baseline methods. Moreover, the experiments were performed on T5 with 48% accuracy, ALBERT with 95% accuracy, RoBERTa + BERT with 95% accuracy for the comparison. Local Interpretable Model-agnostic Explanations (LIME) are applied to enhance interpretability and transparency, identifying key textual features influencing predictions. The results have shown that our hybrid model is not only more accurate than state-of-the-art methods but also provides human-understandable insights, which is a more transparent and effective solution for detecting misinformation on social media.