Anonymizer Rewrite: Inference-Aware Text Anonymization

Text anonymization often removes explicit identifiers while leaving contextual clues that enable re-identification. We introduce Anonymizer Rewrite, an inference-time pipeline combining explicit entity detection with open-ended discovery of latent risks, context-sensitive protection, full-text rewriting, and evaluator-guided repair. Rewrite separately evaluates residual privacy leakage and semantic utility using question--answer harnesses. Evaluated on RAT-Bench against eight baselines, Rewrite achieves the lowest re-identification rate across all three difficulty levels—explicit (2%), obfuscated (5%), and latent (3%)—while maintaining strong semantic utility.