Abstract <p>In the current online society, the proliferation of fake news has become a major problem, posing a challenge to the public’s judgment and the credibility of social media. There is an urgent need for effective solutions to identify and filter false information. A multimodal fake news detection model based on adaptive graph convolutional networks was proposed, which combined text information, visual information, and related knowledge concepts for automatic detection of fake news. This model included a knowledge distillation process based on entity connections and point mutual information, as well as a mechanism for extracting initial word vectors and using adaptive graph convolutional networks for feature aggregation. Experiments on two public datasets, PHEME and WEIBO, had confirmed that the accuracy and precision of the proposed model reached 0.9121 and 0.9161 in the WEIBO dataset, and 0.8854 and 0.8848 in the PHEME dataset, respectively. Its recall and F1-score reached 0.9111 and 0.9117 in the WEIBO dataset, and 0.8854 and 0.8851 in the PHEME dataset, respectively. These results were superior to the comparative models in the comparative experiments. The proposed model has achieved excellent performance in fake news detection through adaptive graph convolutional networks and multimodal analysis. Especially when dealing with social media content with complex semantics and formatted information, it demonstrates good robustness and high accuracy.</p>