Repurpose Image Identification for Fake News Detection.

The Internet has become the major channel where users gather and disseminate news and information, and has given rise to problems such as fake news and disinformation. This work tackles a class of fake news where visually similar images from past events are purported as visual evidence to exaggerate the severity of a current news event. These images have been repurposed and possess a great affinity to real news, posing a challenge to the fake news detection task. We propose a multi-stage approach that comprises an event type classifier to determine the type of news event, and an image repurpose detector which utilizes a siamese network to detect whether the news is fake and contains a repurposed image. Evaluation on real-world news datasets show that the proposed solution outperforms state-of-the-art methods and is effective in identifying fake news containing repurposed images.