Towards a Human-AI Hybrid for Adversarial Authorship
Jordan Allred, Sadaira Packer, Gerry Vernon Dozier, Sarp Aykent, Alexicia Richardson, Michael C. King · 2020
In this paper, we compare two types of masking methods for Adversarial Authorship. One method is a human-based interactive form of masking (referred to as AuthorCAAT-V) while the second method is hybrid of three state-of-the-art author masking techniques (referred to as AIM-IT). Our results show that the performances of AuthorCAAT-V and AIM-IT are equal to or better than the performances of the three state-of-the-art author masking techniques in reducing the identification rate of four well-known authorship attribution systems (AASs). Furthermore, our results show that the hybridization of AuthorCAAT-V and AIM-IT provides a greater reduction in the identification rate.