Multimodal fake news detection on online social

Abstract

The rapid growth in the amount of fake news on instant messaging platforms and social media is a very serious concern in our society. Moreover, the ubiquitous availability of smartphones, affordable network infrastructure, and the increasing popularity of social media create opportunities for malicious users to widespread such misinformation among di erent people. Often some people plot fake news and share it through instant messaging and social media platforms to inuence elections, initiate a propaganda, spread violence, cause riots or humiliate others. Fake news is usually created by manipulating images, text, audio, and videos. Existing fake news detection methods focus on solving this issue by taking into account only one modality (i.e., text, images, audio, or videos). However, one modality is not enough to address such a complex problem. Aimed at detecting the fake news automatically, we introduce a fake news mitigation system. It exploits both text and visual features. Speci cally, we extract features from title, content, and the top-image of an article. The model is able to achieve an F1-score of 0.968 on the FakeNews dataset. Experimental results con rm that the fake news mitigation model is useful in detecting fake news on social media and instant messaging platforms. We also propose the second version of the fake news detection model mitigationV2, which also pays special attention on the types of named entities and their relation present in the news, and improving the image only feature space by bringing image tampering into the fold.