Research On Fake News Detection Method Based On Multimodality And Domain Adaptatio

The proliferation of fake news on social media has become an issue that cannot be ignored,and its harmfulness exceeds people’s expectations.With the development of network technology,the forms of fake news are constantly changing.Nowadays,fake news no longer relies solely on textual content for dissemination,but also uses more persuasive multimedia content such as images,videos,and audio to attract and mislead readers,which also brings new opportunities for detecting fake news.This article focuses on fake news content composed of text and images.Traditional fake news detection methods treat it as a natural language processing problem,focusing more on textual features and ignoring the complementary information of other modal data.Most detection methods are only effective in specific news domains,and the trained models lack good generalization ability,which makes it difficult to apply the models to real-world environments.To address these issues,this article proposes three fake news detection methods,with the main contributions as follows:1.A feature fusion-based multimodal fake news detection method FND is proposed.Specifically,Bi LSTM and Res Net50 are used to extract text features and image features from news data,and the obtained features are weighted using a soft attention mechanism.After passing through the attention layer,the text features and image features are merged into a "text-image" fusion feature.Finally,we used a multilayer perceptron to perform binary classification on the multimodal features,achieving a multimodal fake news detector based on feature fusion.A large number of experiments on real-world datasets show that the feature fusion-based multimodal fake news detection method outperforms traditional detection methods based on either text features or image features.2.A multi-modal fake news detection method FNM based on Siamese networks was proposed.Specifically,FNM drew on the idea of Siamese networks and used the phenomenon of semantic inconsistency between text and image in fake news.After extracting the features of the two modalities,FNM mapped the features from the original space to a new target space through a matching sub-network,so that the simple distance between features in the target space was close to the semantic distance between features in the original space.The matching sub-network was optimized by minimizing the contrastive loss to measure the degree of semantic matching between text and image,and finally output the prediction value of the news’ authenticity.Experimental results show that FNM outperforms other methods on different datasets and evaluation metrics,demonstrating its feasibility and effectiveness in practical applications.3.A domain adaptive fake news detection method FND+Filter and FNM+Filter based on domain adaptation are proposed.Specifically,the domain filter structure proposed in this chapter applies the idea of generative adversarial networks and establishes a minimax game between the multimodal feature extractor and the news domain classifier.By introducing a gradient reversal layer,while ensuring the decrease of the loss of the news authenticity classifier,the loss of the news domain classifier is maximized as much as possible,so that the domain information in the multimodal features can be filtered,the feature difference between the source domain and the target domain can be weakened,and the generalization ability of the model can be improved,achieving the domain adaptation of the fake news detection model.