Poster: Adversarial Examples for Hate Speech Classifiers

With the advent of the Internet, social media platforms have become an increasingly popular medium of communication for people. Platforms like Twitter and Quora allow people to express their opinions on a large scale. These platforms are, however, plagued by the problem of hate speech and toxic content. Such content is generally sexist, homophobic or racist. Automatic text classification can filter out toxic content so some extent. In this paper, we discuss the adversarial attacks on hate speech classifiers. We demonstrate that by changing the text slightly, a classifier can be fooled to misclassifying a toxic comment as acceptable. We attack hate speech classifiers with known attacks as well as introduce four new attacks. We find that our method can degrade the performance of a Random Forest classifier by 20%. We hope that our work sheds light on the vulnerabilities of text classifiers, and opens doors for further research on this topic.