Preprint
Avengers Ensemble! Improving Transferability of Authorship Obfuscation
ArXiv.org
09/15/2021
Abstract
Stylometric approaches have been shown to be quite effective for real-world authorship attribution. To mitigate the privacy threat posed by authorship attribution, researchers have proposed automated authorship obfuscation approaches that aim to conceal the stylometric artefacts that give away the identity of an anonymous document's author. Recent work has focused on authorship obfuscation approaches that rely on black-box access to an attribution classifier to evade attribution while preserving semantics. However, to be useful under a realistic threat model, it is important that these obfuscation approaches work well even when the adversary's attribution classifier is different from the one used internally by the obfuscator. Unfortunately, existing authorship obfuscation approaches do not transfer well to unseen attribution classifiers. In this paper, we propose an ensemble-based approach for transferable authorship obfuscation. Our experiments show that if an obfuscator can evade an ensemble attribution classifier, which is based on multiple base attribution classifiers, it is more likely to transfer to different attribution classifiers. Our analysis shows that ensemble-based authorship obfuscation achieves better transferability because it combines the knowledge from each of the base attribution classifiers by essentially averaging their decision boundaries.
Details
- Title: Subtitle
- Avengers Ensemble! Improving Transferability of Authorship Obfuscation
- Creators
- Muhammad Haroon - University of California, DavisMuhammad Fareed Zaffar - Lahore University of Management SciencesPadmini Srinivasan - University of IowaZubair Shafiq - University of California, Davis
- Resource Type
- Preprint
- Publication Details
- ArXiv.org
- ISSN
- 2331-8422
- Number of pages
- 13 pages
- Comment
- Submitted to PETS 2021
- Language
- English
- Date posted
- 09/15/2021
- Academic Unit
- Nursing; Computer Science; Business Analytics
- Record Identifier
- 9984161432102771
Metrics
18 Record Views