IGWO-MRC: optimization-guided multimodal fake news detection via residual collaboration

As social media becomes pervasive and content generation technologies advance rapidly, fake news increasingly employs cross-modal camouflage strategies, significantly amplifying its spread and perceptual deception in online environments. Traditional unimodal detection methods exhibit limited generalizability and representational capacity when confronted with modality heterogeneity, semantic inconsistency, and insufficient cross-modal interaction modeling. Therefore, this paper proposes a novel multimodal fake news detection framework, IGWO-MRC. First, an Improved Grey Wolf Optimizer (IGWO) performs efficient feature selection through Tent chaotic initialization, dynamic position updating, a dual-objective adaptive fitness function, and hybrid perturbation mechanisms, enhancing search diversity, convergence stability, and feature discriminativeness. PCA further compresses the selected features into compact representations. We incorporate a Modality-Difference Guided Feature Alignment (MDFA) module and a Residual Gated Fusion (RGF) mechanism to enhance semantic alignment and fusion stability, balancing inter-modal coupling and residual regulation. Additionally, a Squeeze-and-Excitation (SE) block reweights fused features based on channel importance, thereby improving discriminative capacity. Finally, a joint loss of Contrastive Modality Clustering Loss (CM-CL) and Focal Loss is employed to sharpen class boundaries and reinforce robustness under noisy or imbalanced conditions. These components collectively ensure accurate, generalizable, and robust multimodal fake news detection. Extensive experiments conducted on a large-scale Chinese multimodal fake news dataset demonstrate that the proposed IGWO-MRC framework achieves an average accuracy improvement of 10.13% over baseline models, confirming its effectiveness, robustness, and adaptability in complex cross-modal semantic scenarios.