No "Love" Lost: Defending Hate Speech Detection Models Against Adversaries

Although current state-of-the-art hate speech detection models achieve praiseworthy results, these models have shown themselves to be vulnerable to attacks. Easy-to-execute lexical evasion schemes such as removal of whitespace from a given text creates significant issues for word-based hate speech detection models. In this paper, we reproduce the results of five cutting-edge models as well as four significant evasion schemes from prior work. These schemes are required to maintain readability which enables us to recreate the original data. We present several new defenses that leverage this need for maintained meaning and readability, and these schemes perform on par with or exceed the results of adversarial retraining. Furthermore, we demonstrate that each lexical attack or evasion scheme can be overcome with our new defense mechanisms with some reducing the effectiveness of the scheme to a mere .1 to .01 drop in F-1 score. We also propose a new evasion scheme that outperforms those in previous work along with a corresponding defense. Using our results as a foundation, we contend that hate speech detection models can be defended against lexically morphed data without the need for significant retraining. Our work suggests that by utilizing the requirement for preserved meaning, one can create a suitable defense against evasion schemes with a high reversal rate.

I. Introduction

Armed with the power of anonymity, many groups and individuals abuse social media tools in order to proliferate messages advocating for violence and suppression of another people. Such language and its spread across online social networks (OSNs) presents several issues to the managers of such networks, but there is an inherent issue with rooting out such speech. The definition and classification of hate speech has not reached a universal consensus. Merriam-Webster designates hate speech as language "expressing hatred of a particular group of people" while the European Union defined it as speech that "spread, incite, promote or justify racial hatred, xenophobia, antisemitism or other forms of hatred based on intolerance". There is a notable distinction between the two definitions above; the latter places an emphasis on hateful "calls to action" while the former does not. This difference, the "call to action", typically is the borderline between merely offensive speech and hateful speech in the eyes of OSN moderators and admins. Regardless, the identification of hate speech a persistent problem for OSN moderators, and due to the sheer volume of generated content on said networks, manual moderation is infeasible.