Fake News Detection Using Transfer Learning with Strategic Layer Freezing on a Fine-Tuned RoBERTa Model

Abstract

The rapid proliferation of misinformation across digital platforms has made automated fake news detection a critical challenge in natural language processing. This paper presents a transfer learning approach with strategic layer freezing for fake news detection, applied to a pretrained RoBERTa-based classifier. We freeze the lower six encoder layers (layers 0–5) of the pretrained model to preserve general-purpose linguistic representations, while retraining the upper six layers (layers 6–11) and the classification head on the GonzaloA/Fake News dataset comprising 40,587 labeled English-language news articles. Our approach achieves 98.19% accuracy on the held-out test set with an F1-score of 0.98, outperforming both the unmodified pretrained model (89.1%) and full fine-tuning of all layers (96.4%) by significant margins. We employ a cosine learning-rate scheduler with AdamW optimization and early stopping to prevent catastrophic forgetting of pretrained knowledge. Analysis of the confusion matrix on 2,434 test samples reveals only 44 total misclassifications (20 false negatives and 24 false positives), confirming strong performance across both classes. The fine-tuned model has been publicly deployed on the Hugging Face Hub for community use and reproducibility.