The proliferation of artificial-intelligence-generated fake content demands robust detection technologies. Traditional unimodal approaches fail to capture cross-modal dependencies, while existing multimodal methods suffer from non-stochastic latent representations that limit nuanced interaction modeling. To address this limitation, we propose SBSD-Detector, a novel framework based on stick-breaking smoothed-Dirichlet distributions for probabilistic multimodal latent representation learning. The proposed model explicitly captures uncertainty and cross-modal interactions within a unified deep learning architecture. We evaluate SBSD-Detector on three benchmark datasets (Twitter, Weibo, and Fakeddit), where it consistently outperforms strong multimodal baselines in terms of accuracy and F1-score. These results demonstrate the effectiveness of probabilistic latent-variable modeling for improving robustness and generalization in multimodal fake-news detection.
Ojo et al. (2026) studied this question.