I Am No One: Style-Aware Paraphrasing for Text Anonymization
Authorship attribution models can re-identify users from anonymized text by exploiting stable stylistic fingerprints, creating privacy risks even after explicit identifiers are removed The proposed approach uses pretrained large language models to build compact stylistic profiles from minimal samples and rewrites text to suppress identifiable style markers while preserving meaning The method reduces authorship attribution F1 by 60-70% across blog and review datasets while maintaining content qua
Analysis
TL;DR
- Authorship attribution models can re-identify users from anonymized text by exploiting stable stylistic fingerprints, creating privacy risks even after explicit identifiers are removed
- The proposed approach uses pretrained large language models to build compact stylistic profiles from minimal samples and rewrites text to suppress identifiable style markers while preserving meaning
- The method reduces authorship attribution F1 by 60-70% across blog and review datasets while maintaining content quality and readability
- This approach substantially outperforms both differential privacy-based and non-DP baseline anonymization methods
- The risk extends to speech-derived text like ASR transcripts, where stylometric leakage persists even after acoustic anonymization
Why It Matters
This research addresses a critical and often overlooked privacy vulnerability: even when personal identifiers are stripped from text, stylistic fingerprints can still re-identify authors, posing significant risks for content publishing, analytics, and speech-derived text applications. For AI practitioners building text processing pipelines, this highlights the need to consider stylometric re-identification as a privacy threat alongside traditional PII removal.
Technical Details
- The approach is prompt-driven and style-aware, leveraging pretrained large language models to construct compact stylistic profiles from minimal text samples
- It rewrites anonymized text to suppress identifiable stylistic markers while preserving semantic meaning and readability
- Evaluated across blog and review datasets, measuring performance via authorship attribution F1 scores
- Compared against differential privacy-based anonymization methods and non-DP baselines, demonstrating superior utility-privacy tradeoffs
- The method is applicable to both written text and ASR transcripts from meetings and call-center conversations
Industry Insight
- Organizations publishing user-generated content or analytics should implement style-aware anonymization as a defense against authorship re-identification, not just PII redaction
- The 60-70% F1 reduction demonstrates that LLM-based paraphrasing can achieve strong privacy guarantees without the severe quality degradation typical of differential privacy approaches
- As speech AI systems increasingly transcribe conversations, stylometric leakage in ASR outputs represents an emerging privacy vector that warrants attention in compliance and data governance frameworks
Disclaimer: The above content is generated by AI and is for reference only.