Lightweight Multimodal Frameworks for Fake News Detection Using Cross-Modal Consistency Modeling and Parameter-Efficient Attention

The rapid proliferation of fake news poses significant societal and psychological risks, driving an urgent need for automated detection systems. While conventional unimodal methods fail to capture complex cross-modal relationships, existing multimodal frameworks are often computationally prohibitive for real-world deployment. To address this limitation, this paper proposes two lightweight, high-performance multimodal frameworks for fake news detection: Cross-Modal Consistency Modeling (CMCM) and Parameter-Efficient Cross-Modal Attention (PE-CMA). The CMCM architecture utilizes a similarity branch to explicitly model semantic alignment and misalignment between text–image pairs, effectively identifying mismatched content characteristic of fake news. Conversely, the PE-CMA architecture introduces a lightweight cross-modal attention mechanism combined with a learned token pooling technique. This compresses token representations into a highly informative subset prior to attention application, significantly lowering computational complexity while preserving cross-modal interactions. Both frameworks leverage MobileNetV3-Large and TinyBERT as highly efficient image and text encoders, respectively. Extensive evaluations on the Fakeddit and Weibo datasets demonstrate the superiority of the proposed methods. On Fakeddit, CMCM and PE-CMA achieve classification accuracies of 92.29% and 91.17%, respectively. On Weibo, PE-CMA and CMCM deliver accuracies of 91.81% and 91.06%, respectively. Ultimately, these architectures outperform several state-of-the-art baselines, offering an optimized balance between classification accuracy and computational efficiency.