The paper introduces ALISON, a method for authorship obfuscation that preserves the semantics of a text while preventing authorship attribution models from correctly identifying the author. This method is a response to privacy concerns raised by state-of-the-art authorship attribution methods. ALISON is shown to be faster and more successful in obfuscation than other methods, with the ability to attack three transformer-based authorship attribution methods on two benchmark datasets. The method does not require direct signals from a target authorship attribution classifier during obfuscation and utilizes unique stylometric features for explainable obfuscation.
Publication date: 2 Feb 2024
Project Page: https://github.com/EricX003/ALISON
Paper: https://arxiv.org/pdf/2402.00835
