Adversarial Arabic Fake News Detection Based on Machine Learning

Fake news has become a major issue owing to its quick growth on the internet, difficulty in identifying it from true news, and people's reliance on social media platforms as primary sources. It has negative effects at several levels, including individual, communal, political, and economical. Detecting Arabic fake news requires significant effort owing to limited datasets and studies in the sector. In this study, Employ Perturbation Adversarial attacks was applied as a regularization technique for fake news classification. Adversarial examples are generated by perturbing the model's word embedding matrix. The AraBERTv2 model is utilized for preprocessing operations. Five different machine learning models were trained on clean data and tested using both clean and adversarial examples to evaluate the generalization capabilities of the classification models. To address the scarcity of Arabic datasets, A translated English fake news was utilized dataset Experimental results indicate that the LightGBM and AdaBoost algorithms exhibited the best performance (90%) compared to other classifiers.