GAConF: A Multimodal Fake News Detection Framework Fusing Gated Attention and Category Perception Losses

Fake news on social media is often disseminated through a combination of text and images, introducing challenges due to modality inconsistency and structural heterogeneity. To address these issues, a novel multimodal fake news detection method, named GAConF (Gated Attention with Category-aware Fusion), is proposed. The method employs a language-adaptive text encoder—MacBERT for Chinese and BERTweet-base for English—to extract contextual textual features, and uses a vision encoder (ViT) to obtain semantic representations of images. A gated attention fusion module is designed to perform dynamic interaction and selective fusion across modalities, enhancing multimodal feature alignment. To mitigate the effect of class imbalance during training, a category-weighted Focal Loss is adopted to optimize the classification task. Experiments conducted on real-world Chinese and English multimodal social media datasets demonstrate that the proposed approach outperforms existing state-of-the-art methods in terms of accuracy and F1-score, showing superior generalization and robustness.