Style-aware paraphrasing cuts authorship attribution F1 by 60-70% while preserving meaning

chaumian · x · 2026-09-14

A paper accepted at Interspeech 2026 tackles stylometric re-identification: authorship attribution models can deanonymize text via stable style fingerprints even after identifiers are removed, and DP-based anonymization severely degrades quality. The proposed style-aware, prompt-driven method uses pretrained LLMs to build compact stylistic profiles and rewrite text to suppress identifiable style markers. On blog and review datasets it reduces attribution F1 by 60-70% while maintaining readability, substantially outperforming DP and non-DP baselines. Risk also extends to ASR transcripts of meetings and call centers.

Original post →

More from Safety

Safety channel →