A Multi-Modal Iterative Guided Framework for Fake News Detection

In recent years, multi-modal fake news detection methods have been widely used and have shown good detection performance in identifying common fake news scenarios, i.e., different tweets are accompanied by different images. However, many fake news stories involve multiple tweets using the same image. Existing detection methods do not fully consider this feature, resulting in unsatisfactory performance in this case. To address this omission problem, we proposed a Multi-modal Iterative Guided Framework (MIGF), where the entire detection process is completed in the latent space. Specifically, first, we construct a latent space at the feature level to alleviate the problem of inconsistent multi-modal feature space and distribution. Then, at the multi-feature interaction level, a multi-modal cross-guided attention mechanism is designed for the interaction of text and visual features. Finally, in terms of multi-view feature extraction and learning, an iterative guided mixture-of-expert system is used to guide the extraction of multi-modal features. Comparative experiments show that the comprehensive performance of MIGF surpasses the baseline models, verifying its effectiveness and superiority. Ablation experiments further quantify the contribution of each module, proving that the overall architecture is indispensable.