Toward Deployable Disinformation Defense: Benchmarking Lightweight Transformers on FakeNewsNet

Disinformation poses a persistent threat to news integrity and online trust, spreading six times faster than factual information on social media platforms. While some transformer-based models achieve state-of-the-art performance in fake news detection, their high computational costs and lack of interpretability limit deployment capabilities and operational transparency. This work benchmarks three lightweight transformer models on FakeNewsNet datasets to address critical gaps in scalability, interpretability, and applicability for fake news detection. Results show that DistilBERT achieves the highest accuracy, precision, F1-score, and AUC, while ALBERT demonstrates superior recall. TinyBERT offers the fastest inference time with minimal memory footprint, making it ideal for edge deployment. Explainability analysis using SHAP and LIME reveals that all models identify similar semantic patterns as primary indicators of fake news, with model-specific attention mechanisms reflecting their respective compression strategies. This study supports the development of scalable, transparent, and high-performing disinformation detection systems suitable for real-world deployment where efficiency, interpretability, and reliability are essential.