This research addresses the risk of re-identification through stylistic fingerprints in anonymized text. Authorship attribution models can identify users even after removing explicit identifiers. This problem extends to speech-derived text, such as transcripts from meetings and call centers. Existing differential privacy-based anonymization techniques often degrade text quality. The paper proposes a style-aware paraphrasing method. This method uses pretrained large language models to create compact stylistic profiles from small samples of text. The system then rewrites the text to remove identifiable style markers while maintaining the original meaning. Experiments on blog and review datasets demonstrate a significant reduction in authorship attribution F1 scores, achieving a 60-70% decrease compared to differential privacy and non-differential privacy baselines. The approach maintains content quality and readability. The research focuses on prompt-driven anonymization. The system leverages the capabilities of large language models for this task.
Source: https://arxiv.org/abs/2609.12341