Arabic Fake News Detection Using a Hybird Arabert and 2-D CNN Architecture

Abstract

The rapid proliferation of disinformation on social media platforms presents a significant challenge to the integrity of digital information, particularly within the Arabic-speaking world. This paper proposes a hybrid deep learning architecture, AraBERT-2DCNN, for the

automated detection of fake news using the Arabic Fake News Dataset

(AFND). The proposed approach integrates the contextual representation capabilities of the AraBERT transformer with a multiscale two-dimensional Convolutional Neural Network (2D-CNN) to capture both global semantic information and local linguistic patterns. To address common challenges such as overfitting and catastrophic forgetting in transformer-based models, a two-stage fine-tuning strategy is employed, along with Focal Loss to mitigate class imbalance. Experimental results demonstrate that the proposed model achieves a classification accuracy of 90.45%, with F1-scores of 0.93 and 0.86 for credible and non-credible news, respectively. These findings indicate that stabilizing transformer training and integrating convolutional feature extraction significantly enhance generalization performance across diverse news domains.