Personalized Author Obfuscation with Large Language Models

TL;DR

The paper investigates whether large language models can obfuscate author identity through paraphrasing, revealing substantial variability across authors with a bimodal distribution of obfuscation effectiveness. It compares zero-shot prompting to a SHAP-informed personalized prompting approach, demonstrating that author-specific prompts can reduce attribution more reliably across diverse datasets and AV models, including BERT and LR-based verifiers. By analyzing SHAP-valued features and employing Hartigan's Dip Test, the study highlights the multimodal nature of obfuscation success and offers a practical path to mitigate it. The work contributes to privacy-aware text synthesis and informs design choices for user-tailored obfuscation strategies, while acknowledging limitations such as dataset size and domain-specific effects.

Abstract

In this paper, we investigate the efficacy of large language models (LLMs) in obfuscating authorship by paraphrasing and altering writing styles. Rather than adopting a holistic approach that evaluates performance across the entire dataset, we focus on user-wise performance to analyze how obfuscation effectiveness varies across individual authors. While LLMs are generally effective, we observe a bimodal distribution of efficacy, with performance varying significantly across users. To address this, we propose a personalized prompting method that outperforms standard prompting techniques and partially mitigates the bimodality issue.

Figures (1)

• Figure 1: Top features with highest average SHAP values for a given user. The side with higher concentration of red dots indicate the affect of increasing feature value on model's prediction.