FDNN: A Feature-Driven Neural Network Framework for Automated Fake News Detection

The rapid spread of fake news across social media platforms has emerged as a critical issue, shaping public opinion and amplifying misinformation on a global scale. This challenge necessitates the development of robust, automated systems capable of accurately detecting misleading content. While traditional machine learning approaches such as Logistic Regression (LR) are lightweight and straightforward to implement, they rely primarily on shallow textual features, limiting their ability to capture deeper semantic patterns. In contrast, transformer-based models like BERT and its optimized successor RoBERTa excel at contextual language understanding, enabling more accurate classification. To harness the strengths of both transformer architectures, we propose a hybrid of BERT and RoBERTa as HBERoB model, where predictions from both models are strategically combined to improve accuracy, generalization, and robustness. Using a labeled dataset of news articles, we applied preprocessing steps including text cleaning, tokenization, and transformation before evaluating four models: LR, BERT, RoBERTa, and the proposed HBERoB. Both transformer models were fine-tuned through transfer learning to adapt to the dataset effectively. Model performance was assessed using accuracy, precision, recall, and F1-score. Experimental results show that the HBERoB model consistently outperforms individual models, with notable improvements in precision and recall while maintaining efficiency. This work is limited to English textual data and does not extend to multimedia or real-time detection, which are suggested for future exploration.