D-FEND: A Diffusion-Based Fake News Detection Framework for News Articles Related to COVID-19

Abstract

The social confusion caused by the recent pandemic of COVID-19 has been further facilitated by fake news diffused via social media on the Internet. For this reason, many studies have been proposed to detect fake news as early as possible. The content-based detection methods consider the difference between the contents of true and fake news articles. However, they suffer from the two serious limitations: (1) the publisher can manipulate the content of a news article easily, and (2) the content depends upon the language, with which the article is written. To overcome these limitations, the diffusion-based fake news detection methods have been proposed. The diffusion-based methods consider the difference among the diffusion patterns of true and fake news articles on social media. Despite its success, however, the lack of the diffusion information regarding to the COVID-19 related fake news prevents from studying the diffusion-based fake news detection methods. Therefore, for overcoming the limitation, we propose a diffusion-based fake news detection framework (D-FEND), which consists of four components: (C1) diffusion data collection, (C2) analysis of the data and feature extraction, (C3) model training, and (C4) inference. Our work contributes to the effort to mitigate the risk of infodemics during a pandemic by (1) building a new diffusion dataset, named CoAID+, (2) identifying and addressing the class imbalance problem of CoAID+, and (3) demonstrating that D-FEND successfully detects fake news articles with 88.89% model accuracy on average.