Fine-Grained Differences-Similarities Enhancement Network for Multimodal Fake News Detection

The use of social media has proliferated dramatically in recent years due to its increasing reach and ease of use. Along with this enlarged influence of social media platforms and the relative anonymity afforded to content contributors, an increasingly significant proportion of social media is composed of untruthful or “fake” news. Hence for various reasons of personal and national security, it is essential to be able to identify and eliminate fake news sources. The automated detection of fake news is complicated by the fact that most news posts on social media takes very diverse forms, including text, images, and videos. Most existing multimodal fake news detection models are structurally complex and not interpretable; the main reason for this is the difficulty of identifying essential features which characterize fake social media posts, leading to different models focusing on multiple different aspects of the news detection task. In this paper, we show that contrasting the different and similar (DS) features of social media posts serves as an important identifying marker for their authenticity, with the consequence that we only need to direct our attention to this aspect when designing a multimodal fake news detector. To address this challenge, we propose the Fine-Grained Differences-Similarities Enhancement Network (FG-DSEN), which improves detection with a simple and interpretable structure to enhance the DS aspect between images and text. Our proposed method was evaluated on two different language social media datasets, Weibo in Chinese and Twitter in English. It achieved accuracies 3% and 3.8% higher than other state-of-the-art methods, respectively.