When we share text data—whether it is clinical notes, social media posts, or forum discussions—we risk exposing sensitive information. This exposure happens not just through explicit identifiers like names and dates, but also through contextually embedded details and implicit linguistic fingerprints (stylometry). As Gary Kacmarcik et al. demonstrated in Obfuscating Document Stylometry to Preserve Author Anonymity, even minor targeted edits to a small number of words can drastically alter an author's linguistic fingerprint.
However, traditional sanitization methods like aggressive redaction or binary masking often destroy the semantic value of the text, rendering it useless for downstream analysis, as noted by Prasadi Abeywardana et al. in A Privacy Preserving Data Publishing Middleware for Unstructured, Textual Social Media Data and by Ildikó Pilán et al. in The Text Anonymization Benchmark (TAB): A Dedicated Corpus and Evaluation Framework for Text Anonymization. The core challenge of privacy-aware text rewriting is navigating this strict trade-off: how do we obscure identifying information while preserving the text's original meaning, fluency, and utility?
Why Standard Differential Privacy Fails on Text
To understand why this is difficult, consider Local Differential Privacy (LDP), the mathematical gold standard for data protection. Standard LDP requires that any input token has a non-zero probability of mapping to any other token in the vocabulary.
In a discrete text space, this treats all words as equidistant; the word "cat" is mathematically as close to "feline" as it is to "refrigerator." Forcing this uniform indistinguishability injects massive noise, transforming coherent sentences into grammatically disconnected "word salad," as argued by Fengran Mo et al. in A Customized Text Sanitization Mechanism with Differential Privacy. As Stephen Meisenbacher et al. observed in Thinking Outside of the Differential Privacy Box: A Case Study in Text Privatization with Language Model Prompting, strict differential privacy constraints frequently preserve high-level semantic meaning at the expense of linguistic coherence, imposing a heavy readability tax.
The Continuous Space Escape: Metric Differential Privacy
To preserve meaning, a privacy mechanism must account for the fact that some words are semantically closer than others. Oluwaseyi Feyisetan et al. addressed this in Calibrating Mechanisms for Privacy Preserving Text Analysis by introducing Metric Differential Privacy (mDP) to text processing.
Instead of operating on discrete tokens, mDP maps words into a continuous vector space (embeddings) where geometric distance correlates with semantic similarity. By adding calibrated noise directly to these vectors and mapping the perturbed vector back to its nearest vocabulary neighbor, the mechanism swaps words for their synonyms or contextually similar alternatives.
[Original Token: "doctor"] ───> Embedding Space ───> [Add Calibrated Noise] ───> [Nearest Neighbor: "physician"]
In Privacy- and Utility-Preserving Textual Analysis via Calibrated Multivariate Perturbations, the authors demonstrated that this geometric approach maintains enough semantic utility for downstream machine learning models to remain highly accurate. To prevent these perturbations from drifting into complete incoherence, Di Wang et al. developed a High-Dimensional Truncated Laplacian mechanism in Private Language Models via Truncated Laplacian Mechanism, which mathematically bounds the noise to ensure the output remains fluent and logically consistent.
Generative and Adversarial Rewriting
While token-swapping preserves high-level semantics for classification tasks, it struggles with multi-word entities and complex syntax. To maintain natural human structure, researchers have pivoted toward generative models that rewrite text at the sentence or document level.
Haohan Bo et al. introduced an Embedding Reward Auto-Encoder (ER-AE) in ER-AE: Differentially Private Text Generation for Authorship Anonymization, combining a differential privacy mechanism with reinforcement learning to optimize directly for readability and authorship obfuscation.
More recently, Large Language Models (LLMs) have been deployed in feedback-guided adversarial frameworks. In Language Models are Advanced Anonymizers (and expanded in Language Models are Advanced Anonymizers), Robin Staab et al. designed a system where one LLM acts as a detective trying to uncover private details, while another acts as an editor, iteratively refining the text to block those specific inferences without destroying the story.
┌────────────────────────────────────────────────────────┐
│ │
│ ┌─────────────────┐ ┌───────────────────┐ │
│ │ │ Rewrite │ │ │
│ │ LLM Editor ├─────────>│ Privatized Text │ │
│ │ │ │ │ │
│ └────────▲────────┘ └─────────┬─────────┘ │
│ │ │ │
│ │ Feedback │ Inspect │
│ │ │ │
│ ┌────────┴────────┐ ┌─────────▼─────────┐ │
│ │ │ │ │ │
│ │ LLM Detective │<─────────┤ Private Inferences│ │
│ │ │ Expose │ │ │
│ └─────────────────┘ └───────────────────┘ │
└────────────────────────────────────────────────────────┘
To eliminate the privacy risk of sending sensitive data to external cloud APIs during this process, Kyuyoung Kim et al. introduced the SEAL framework in Self-Refining Language Model Anonymizers via Adversarial Distillation, which distills both the anonymizing and critiquing capabilities into smaller, local models. To navigate the massive search space of possible rewrites without expensive fine-tuning, Shuo Huang et al. proposed NaPaRe in Zero-Shot Privacy-Aware Text Rewriting via Iterative Tree Search, using iterative tree search guided by a reward model to explicitly balance privacy, utility, and linguistic naturalness.
Obfuscating Stylistic Fingerprints
Privacy protection isn't just about hiding explicit facts; it also requires masking an author's unique writing style. As Chris Emmery notes in Style Obfuscation by Invariance, a fundamental tension exists between altering stylistic markers and maintaining grammatical coherence.
To address this, Janek Bevendorff et al. framed style obfuscation as a heuristic search problem in Heuristic Authorship Obfuscation, strategically swapping characters and words to maximize the "style distance" from the original author while minimizing text quality degradation.
Alternatively, David Ifeoluwa Adelani et al. demonstrated in Preventing Author Profiling through Zero-Shot Multilingual Back-Translation that multilingual back-translation (e.g., translating text from English to a pivot language and back) acts as an effective, zero-shot obfuscator. This process naturally smooths out demographic stylistic markers while retaining over 95% of the original text's semantic utility.
Key Foundational Work
Obfuscating Document Stylometry to Preserve Author Anonymity (2006) — contributed 1 concept elements to this direction.
Author Masking through Translation (2016) — contributed 1 concept elements to this direction.