Arabic Fake News Detection Using Genetically Optimized Deep Learning Models

Abstract

In recent years, the proliferation of fake news across social media and digital platforms has significantly undermined public trust in traditional media outlets, raising serious challenges at the social, political, and health levels. Within the Arabic context, this problem is further exacerbated by the scarcity of effective tools for Arabic language processing. This study proposes a hybrid model that combines AraBERT for extracting deep contextual representations, Convolutional Neural Networks (CNN) for capturing local features, and a Genetic Algorithm (GA) for hyperparameter optimization. The model was evaluated on an Arabic dataset consisting of real and fake news articles, while class imbalance was addressed using class weights. Experimental results show that the GA-optimized AraBERT-CNN hybrid model outperformed both traditional (LSTM) and modern (pure AraBERT) baselines, achieving an overall accuracy more than 96% and higher F1-scores for the minority class (real news). The significance of this study lies in presenting a comprehensive framework that integrates the contextual power of transformer-based models with the adaptive optimization of evolutionary algorithms, thereby contributing to ongoing efforts to combat misinformation in Arabic and to develop more reliable intelligent news verification systems.