Skip to main content

DxG-Fuse: Description-Guided Fusion for Multimodal Fake News Detection via MLLMs

By
Zixuan Yu; Guopu Zhu; Zheming Fan; Xiaolong Li; Xiangyang Luo; Ligang Wu

The rapid spread of fake news on social media poses significant threats to information security and social stability. Traditional multimodal fake news detection methods struggle to effectively incorporate open-world knowledge, limiting their ability to reason about unseen events or entities. Recently, detection methods that directly use multimodal large language models (MLLMs) have emerged, but they suffer from perspective bias, as their judgments tend to rely excessively on certain modalities or reasoning cues, while overlooking complementary evidence from other perspectives. Such bias limits their reliability for fake news detection. To address these challenges, we propose DxG-Fuse, a novel framework for multimodal fake news detection, which is built upon a description-guided fusion paradigm that combines MLLM-generated multi-perspective descriptions with deep feature fusion. Specifically, DxG-Fuse uses an MLLM, guided by an image forgery detector, to generate explanatory descriptions from multiple perspectives, including factual correctness, emotional tone, image-text consistency, and forgery-aware visual cues. These descriptions, along with the original image-text features, are integrated via a deep fusion network equipped with a forgery-description-guided visual attention module. Extensive experiments on three public benchmark datasets demonstrate that DxG-Fuse outperforms state-of-the-art methods in detection accuracy. Ablation studies further confirm the effectiveness of multi-perspective descriptions and the proposed fusion strategy. The source code is available at https://github.com/hityzx/DxG-Fuse

Read on IEEE Xplore