Image-based hoax detection

In the last few years, the impact of information spread through online social networks has continuously grown. For this reason, understanding the trustworthiness of news has become one of the most important challenges for an Internet user, especially during crisis events or in political, health and social issues. As part of a more comprehensive project for the detection of fake news, this paper proposes a machine learning method to evaluate the trustworthiness of a piece of information especially considering its associated image. In the work described in this paper, the training and test datasets have been first collected from the web, downloading more than 1000 images related to trusted and fake Facebook pages. All collected images have been processed using the Google Vision online service for extracting their specific internal details. For each image, various kinds of features have been considered, including its color composition, the recognized objects, the list of sites in which it is published, and eventually the contained text. These details have been then used for training a classifier using different algorithms which allowed us to reach an accuracy of about 85% in hoax identification. Future research will focus on social-network information related to images, to improve the system accuracy and acquire more knowledge about various types of news spread online.