Abstract

The spread of fake news on social networks has become a serious challenge to political and social stability. Previous studies have mainly relied on unimodal textual analyses and have been unreliable in intelligently integrating heterogeneous features. This paper presents a novel method for detecting fake news, called MFND, which is based on the combination of statistical, structural, and semantic features in the form of a deep multimodal architecture. In the proposed model, statistically engineered features, after dimensionality reduction by principal component analysis, are placed alongside features extracted through TF-IDF and GloVe semantic embeddings. The main distinguishing attribute of this research is the use of two parallel convolutional neural networks and a soft attention mechanism layer for dynamic weighting and intelligent integration of these information modalities. Evaluation of the model on two standard datasets GossipCop and Politifact shows the high efficiency of MFND in detecting false news content; In the GossipCop dataset, the proposed strategy achieved 98.12% and 0.9618 in terms of accuracy and F-statistics, respectively, and these results in the Politifact dataset were 95.56% and 0.9418, respectively. These results confirm that integrating statistical features with deep content patterns through the attention mechanism will significantly improve the ability to distinguish authentic news from fake news.