Heuristic Authorship Obfuscation

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1098–1108 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 1098 Heuristic Authorship Obfuscation Janek Bevendorff ∗ Martin Potthast† Matthias Hagen‡ Benno Stein∗ ∗Bauhaus-Universität Weimar †Leipzig University ‡Martin-Luther-Universität Halle-Wittenberg <first>.<last>@uni-{weimar, leipzig}.de <first>.<last>@informatik.uni-halle.de Abstract Authorship verification is the task of deter- mining whether or not two texts were written by the same author. This paper deals with the adversary task, called authorship obfusca- tion: Preventing verification by altering a to- be-obfuscated text. We introduce an approach that (1) models writing style difference as the Jensen-Shannon distance between the charac- ter n-gram distributions of texts, and (2) ma- nipulates an author’s subconsciously encoded writing style in a sophisticated manner using heuristic search. To obfuscate, we explore the huge space of textual variants in order to find a paraphrased version of the to-be-obfuscated text that has a sufficient Jensen-Shannon dis- tance at minimal costs in terms of text quality loss. We analyze, quantify, and illustrate the rationale of this approach, define paraphrasing operators, derive obfuscation thresholds, and develop an effective obfuscation framework. Our authorship obfuscation approach defeats state-of-the-art verification approaches, includ- ing unmasking and compression models, while keeping text changes at a minimum. 1 Introduction Can the authorial style of a text be consistently ma- nipulated? More than a century worth of research on stylometry and authorship analysis could not produce a reliable approach to do so manually. In the context of computational authorship obfusca- tion, a handful of approaches have achieved some limited success but are still rather insufficient. Rule- based approaches are neither flexible, nor is stylom- etry understood well enough to compile rule sets that specifically target author style. Monolingual machine translation-based approaches suffer from a lack of training data, whereas applying multilin- gual translation in a cyclic manner as a workaround has proved to be ineffective. In addition, none of the existing approaches offers a means to control the result quality. Given recent advances in con- trolled text generation, it stands to reason that a lot more can be achieved. In this paper, we depart from the mentioned ob- fuscation paradigms and, for the first time, cast author obfuscation as a heuristic search problem. Given a to-be-obfuscated text, we search for a cost- minimum sequence of tailored paraphrasing op- erations that achieve a significant increase of the text’s style distance to other texts from the same author under a generic writing style representation; costs accrue through operations in terms of their estimated text quality reduction. By designing a hy- brid search strategy that neglects admissibility only in the pooling phase, we obtain a significant reduc- tion of the exponentially growing search space that is induced by the paraphrasing operators, enabling the use of informed search algorithms. Moreover, we developed a sophisticated framework to deal with the conflicting objectives that naturally arise with such kind of complex text synthesis tasks: a compact representation of the search space of para- phrased text variants, and …

Content truncated to 500 words