MAFIA-Net: Multimodal Arabic fake-news identification via hybrid attention networks

Résumé fourni par la source

Today, the rapid proliferation of multimodal fake news poses a serious threat to the integrity of information on social networks. While recent multimodal approaches have advanced the field, existing research predominantly targets English content, leaving Arabic largely underexplored. Moreover, manual annotation of multimodal data remains costly and time-consuming. To address these challenges, we propose a scalable data collection and labeling strategy that leverages existing annotated Arabic unimodal datasets, substantially reducing manual annotation efforts. In addition, we introduce a hybrid attention framework that integrates the MARBERTv2 model for textual representation with EfficientNet-B1 for visual encoding. The architecture incorporates hierarchical cross-modal attention and visual-guided textual attention to effectively capture fine-grained multimodal interactions. Extensive experiments compared to state-of-the-art models demonstrate that the proposed approach consistently outperforms both unimodal and multimodal baselines. Notably, our approach achieves an improvement ranging from 1.72% to 5.64% in the F1-score for the Fake class, and from 1.47% to 3.99% in the Macro-F1 score. The proposed framework achieves an accuracy of 92.45% and a Macro-F1 of 90.34%, establishing a new state-of-the-art for Arabic multimodal fake news detection. The dataset and source code are publicly available at: https://github.com/khairied/MAFIA_Net-Multimodal-Arabic-Fake_news-Identification-via-Hybrid-Attention-Networks .

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.