Fake news detection is growing in importance as a key topic in the information age. However, most current methods rely on pre-trained small language models (SLMs), which face significant limitations in processing news content that requires specialized knowledge, thereby constraining the efficiency of fake news detection. To address these limitations, we propose the FND-LLM Framework, which effectively combines SLMs and LLMs to enhance their complementary strengths and explore the capabilities of LLMs in multimodal fake news detection. The FND-LLM framework integrates the textual feature branch, the visual semantic branch, the visual tampering branch, the co-attention network, the cross-modal feature branch and the large language model branch. The textual feature branch and visual semantic branch are responsible for extracting the textual and visual information of the news content, respectively, while the co-attention network is used to refine the interrelationship between the textual and visual information. The visual tampering branch is responsible for extracting news image tampering features. The cross-modal feature branch enhances inter-modal complementarity through the CLIP model, while the large language model branch utilizes the inference capability of LLMs to provide auxiliary explanation for the detection process. Our experimental results indicate that the FND-LLM framework outperforms existing models, achieving improvements of 0.7%, 6.8% and 1.3% improvements in overall accuracy on Weibo, Gossipcop, and Politifact, respectively.
