Using Machine Learning & AI to tackle Online Hate and 'Anti-X' Sentiment Open Access
Downloadable ContentDownload PDF
Understanding the extent to which machine learning can be used to classify images is a hot topic in artificial intelligence. Applications range from human medicine, to the automatic identification of items in an online marketplace, through to national security issues related to risk assessment. The goal of any image recognition task is to create a machine learning classifier that can correctly identify as many objects pictured in an image as possible. Much of the fundamental research focuses on one particular type of classifier: a Convolutional Neural Network (CNN). A particularly topical area of such image analysis concerns the analysis of human faces. In particular, there is a growing body of work using a CNN to try to identify a person's emotion or gender based on an image of his/her face. This paper examines a novel application of such machine learning in the field of human facial recognition, concerning people's likely membership of online groups that promote hate speech. Specifically, our research uses a CNN to try to determine whether the facial profile picture that individuals choose to represent themselves online, provides an indication of their membership of online social media groups associated with hate speech. Our social media data from online hate groups contains 5,331 profile images containing faces from the VKontakte social media platform ( www.vk.com) and comprises entirely publicly available information. We have no need of, nor do we collect, any private information about individuals. An additional 4,555 faces came from the AffectNet dataset and were used as training data for faces not in an online hate group. We trained a multilayer classifier using the Microsoft Cognitive Toolkit module for Python. Of the total 9,886 images, 6,920 were used as training data, 1,482 were used as test data, and 1,483 were used to evaluate the accuracy of the trained model. Our results show a surprisingly high success rate for the CNN (approximately 70%), in terms of predicting whether a given facial profile belongs to a member of an online hate group. We discuss the likely reasons for this success, as well as providing some initial observations concerning broader 'anti-X' hate aimed at historically marginalized populations, religion, race, women and genders.