DistilBERT-Based Explainable Deceptive and Fabricated Information Detection: Enhancing Accuracy and Transparency in Text Classification

The exponential rise of misinformation and Deceptive and Fabricated Information dissemination across online platforms has significantly impacted political discourse, public health decisions, and social trust. While deep learning techniques, including transformers, have advanced Deceptive and Fabricated Information detection, the interpretability of such models remains inadequate, preventing broader adoption in mission-critical applications. This paper investigates a DistilBERT-based explainable architecture for Deceptive and Fabricated Information detection that balances accuracy, efficiency, and interpretability. DistilBERT, a lightweight version of BERT, offers computational efficiency, making it suitable for real-time applications. We propose an Explainable AI (XAI) integration using SHAP (SHapley Additive Explanations) and LIME (Local Interpretable Model-Agnostic Explanations) to enhance transparency of classifications. Evaluation is conducted on benchmark datasets like LIAR (12.8K labeled claims) and Deceptive and Fabricated Information (political and celebrity Information) using metrics such as Accuracy, Precision, Recall, and F1-score. Our model achieves 97% accuracy and 97% F1-score, outperforming baseline methods such as Logistic Regression, Random Forest, SVM, and LGBM classifiers. Beyond performance, explain ability highlights contextually important tokens for classification, aiding end-users in trusting and understanding model predictions. The results demonstrate the potential of DistilBERT for real-world, high-stakes deployment in misinformation regulation.