Fake News Detection Based on Two-Branch Network and Domain Adversarial

Abstract

Fake news detection is essential for society, however, implicit state information in features is ignored in multimodal fake news detection, resulting in inefficient of feature. There are also poor domain generality of features problems. So, a Two-Branch Network with Domain Adversarial (TBNDA), is proposed. Firstly, a pre-trained language model is used to encode features on textual information, and the hidden layer of word information and sentence information in the features is extracted separately using a two-branch network. Secondly, a pre-trained residual network model is used to encode the image information, and a two-branch network model is used to extract the different hidden layer image feature information. Finally, a domain adversarial network module is constructed to extract generic features between domains. The accuracy of the proposed model is S9.6% and S4.7% on the Weibo dataset and Twitter dataset respectively. The two-branch network improves the feature representation of images and text, and the domain adversarial network extracts features with generality, enhancing the migration performance of the model and improving the detection of fake news.