Gated bidirectional cross-modal attention framework with domain generalization for multimodal fake news detection

Abstract The rapid proliferation of multimodal misinformation on social media necessitates systems that identify fake news across domains. Existing studies rely on simple concatenation-based fusion, exhibiting poor generalization across heterogeneous platforms with varied content patterns and visual semantics. To address this, this study presents Gated Cross Balanced Attention Network (GCBA-Net), a novel framework that enhances fake news detection, robustness and domain adaptability. GCBA-Net employs an adaptive gating to independently weight textual and visual inputs. To assess architectural design choices on detection performance, four variants of GCBA-Net were developed varying in dimensionality, projection strategy, gating function, feed-forward activation, attention head count, and attention type. Domain generalization is achieved through Gradient Reversal Layer (GRL) and Maximum Mean Discrepancy (MMD) techniques. Evaluation on Weibo and Fakeddit datasets demonstrates GCBA-Net variants consistently outperform state-of-the-art baselines. The best variant on the Weibo achieved nearly 0.99 accuracy, improving over the EANN baseline. Performance improved from 0.77 for the existing fusion model to over 0.98 accuracy on Fakeddit, with precision, recall, and F1-score exceeding 0.97. GCBA-Net v1, v3 and v4 showed strong performance, while GCBA-Net v2 yielded moderate results. These findings demonstrate that GCBA-Net is an adaptable, effective framework for detecting multilingual fake news in real-world scenarios.