Dynamic lifecycle induced authenticity analysis for multi-modal fake news detection

The news lifecycle implies the dynamic activity level of news throughout its propagation. This lifecycle pattern differs statistically between true and fake news due to the distinct reader interactions and authentication processes involved. Recognizing these distinctions is crucial for accurate detection of misinformation. However, existing fake news detection models often prioritize extracting features associated with propagation or temporal aspects to complement news content, potentially neglecting the significant impact of news lifecycle dynamics on detection. To this end, we propose a pioneering Dynamic Lifecycle-Induced Multi-modal Fake News Detection model (DLIM-FND) that concurrently discerns news authenticity and tracks the news lifecycle based on multi-modal news content and news retweet sequence. Specifically, DLIM-FND introduces an elaborate multi-modal attention mechanism to capture consistent correlation information across multi-modal news content, adopting modal alignment regularization to ensure consistency between modalities. Subsequently, a time encoding module is designed to obtain implicit semantics and temporal features. A novel news lifecycle construction module is introduced to capture the dynamic nature of news cycles. The module is trained under the supervision of the next retweet time prediction task by utilizing news text, retweet content, and timestamps. By fusing multi-modal content features with dynamic news lifecycle representations, DLIM-FND notably enhances its efficacy in detecting fake news. Extensive experiments on three benchmark datasets reveal that DLIM-FND significantly outperforms state-of-the-art detection models. The results demonstrate the effectiveness of incorporating news lifecycle features into the analysis of news authenticity.