Arabic Fake News Detection Through Multi-View Dialect-Aware Adversarial Fusion (MDAF): A Novel LLM-Enhanced Model Architecture

The proliferation of Arabic fake news has become a challenging issue due to the Arabic language's morphological richness and the scarcity of high-quality data corpora. Moreover, existing research often neglects the linguistic diversity of Arabic, as the focus is more on Modern Standard Arabic. To address these issues, this paper proposes the Multi-View Dialect-Aware Adversarial Fusion (MDAF) model. The proposed framework introduces a novel data augmentation-based strategy that utilizes the Fanar-1-9B Large Language Model (LLM) to generate dialectal and adversarial variants of training samples. Subsequently, for classification, MARBERTv2, a state-of-the-art Arabic transformer, is fine-tuned and modified with a concatenated hidden-layer architecture and Multi-Sample Dropout to capture deep linguistic features. The model is evaluated on the VERAARAB and AFND datasets, and experimental results demonstrate MDAF's robust performance, achieving a cross-dialectal F1-score of 85.48% on the VERA-ARAB dataset, representing a statistically significant 7.52 percentage point improvement over the MSA-only baseline, which further highlights the efficacy of dialect-aware adversarial training in low-resource settings.