The increasing spread of misinformation on social platforms has made it necessary to develop automated and accurate fake news detection systems. Traditional methods that focus solely on textual content are ineffective in dealing with more complex fake news that mimic authentic writing styles. This research introduces a novel multi-modal deep learning architecture for fake news detection that achieves a more comprehensive understanding of the content by intelligently integrating non-textual clues (including statistical and behavioral features) and deep textual features. In the proposed model, news features are extracted from distinct modalities: statistically engineered features and behavioral-communicational characteristics of news publishers, character-level features (to deal with misspellings and words outside the vocabulary), as well as word-level semantic features using the Word-level features (Word2Vec) model. Each of these information modalities is processed by a dedicated deep network to produce compact and rich representations of each one. Then, in an innovative step, these representations are intelligently merged using a Multi-Head Self-Attention (MHSA) mechanism to dynamically determine the weight and importance of each modality. Finally, the fused feature vector is used by a SoftMax classifier to finally detect whether the news is fake or authentic. The evaluation results of the proposed method, which were performed on two valid GossipCop and PolitiFact datasets, demonstrated remarkable efficiency. On the GossipCop dataset, accuracy values of 0.99 and F-measure of 0.98 were achieved by the proposed method, demonstrating the high ability of the model in identifying news accurately and completely. Similarly, on the PolitiFact dataset, values of accuracy of 0.96 and F-measure of 0.95 were acquired. This high performance on both datasets indicates an evident dominance over the comparative approaches and confirms the validity and performance of the proposed multi-modal technique for detecting fake news.
Dang et al. (Thu,) studied this question.