A Survey on Current Trends and Recent Advances in Text Anonymization

Tobias Deußer, Lorenz Sparrenberg, Armin Berger, Max Hahnbück, Christian Bauckhage, Rafet SifaarXiv 2025 · DOI · arXiv

Here is a concise, engaging summary of the survey paper, tailored for a technically literate audience.

Paper Summary: A Survey on Current Trends and Recent Advances in Text Anonymization

TL;DR

This survey maps the rapid evolution of text anonymization from traditional Named Entity Recognition (NER) to the era of Large Language Models (LLMs). It highlights how LLMs act as a double-edged sword—serving as highly sophisticated tools for rewriting sensitive text while simultaneously posing unprecedented threats as potent de-anonymization adversaries.

What Problem Does This Paper Address?

The explosion of textual data—ranging from electronic health records and legal briefs to customer reviews—presents a massive privacy risk under strict regulatory frameworks like GDPR and HIPAA. Traditional text anonymization focused primarily on redacting explicit Personally Identifiable Information (PII) like names and phone numbers.

However, modern analytical tools can easily exploit "quasi-identifiers" (contextual clues that, when combined, uniquely identify an individual) or distinct writing styles to re-identify users. The core challenge the field faces is balancing robust privacy protection against the utility of the data for downstream tasks (like algorithmic trading or medical research) in an ecosystem where attackers have access to powerful AI.

Methods

As a comprehensive survey, this paper reviews a broad spectrum of methodologies rather than introducing a single new algorithm. The authors categorize the technical landscape into four main paradigms:

  • Foundational NER-Driven Approaches: Utilizing sequence labeling models (e.g., BiLSTM-CRF, BERT, RoBERTa) combined with rule-based systems and gazetteers to detect and mask explicit PII.
  • LLM-Based Anonymization: Leveraging models like GPT-4 for zero-shot or few-shot text rewriting, or distilling the knowledge of massive LLMs into smaller, resource-efficient encoder-only models for local deployment.
  • Formal Privacy Models: Integrating mathematical frameworks like Differential Privacy (DP) and $k$-anonymity directly into text rewriting pipelines (e.g., DP-BART, DP-MLM) to provide provable privacy guarantees.
  • Authorship Anonymization (Obfuscation): Employing reinforcement learning and constrained decoding (e.g., JAMDEC) to alter an author's stylistic fingerprint while preserving the underlying semantic meaning of the text.

Datasets

The paper reviews various datasets and benchmarks used across the literature to train and evaluate these systems, notably highlighting:

  • The Text Anonymization Benchmark (TAB): A widely cited, open-source corpus of English court cases featuring comprehensive annotations of personal information that go beyond traditional PII.
  • CoNLL-2003: A classic named entity recognition dataset frequently used to benchmark the baseline detection capabilities of anonymizers.
  • Domain-Specific Corpora: Various datasets comprising Electronic Health Records (EHRs), noisy call-center transcripts, and online student discussion forums.

Note: As a survey paper, it does not introduce a new dataset.

Key Insights and Ideas

  • The LLM Paradox: LLMs possess an unprecedented contextual understanding that allows them to replace sensitive information fluently. However, their vast background knowledge makes them excellent "de-anonymizers," capable of inferring identities from residual, subtly scrubbed text.
  • Shift to Stylometry: Anonymization is moving beyond simple text redaction. Protecting privacy now requires obscuring stylometric features—the unique linguistic patterns an author leaves behind—especially in sensitive contexts like whistleblowing or social media research.
  • The Privacy-Utility Tension: Formal methods like Differential Privacy offer ironclad mathematical guarantees, but they often result in unnatural, less fluent text. Finding the sweet spot between provable privacy and semantic integrity remains a major hurdle.

Key Results

While quantitative metrics vary across the dozens of cited studies, the authors synthesize several high-level conclusions:

  • Transformers Dominate: Transformer architectures (BERT, RoBERTa) consistently outperform traditional machine learning and rule-based systems in specialized domains like healthcare and finance, though cross-institutional generalization remains difficult.
  • Generative Success: Frameworks like DeID-GPT demonstrate high accuracy in zero-shot medical text de-identification, successfully masking Protected Health Information (PHI) while maintaining text structure.
  • Adversarial Risk is Real: Studies simulating automated re-identification attacks show that while current risks are relatively low for well-anonymized documents, the threat vector will scale rapidly as LLM capabilities advance.

Limitations and Future Work

The authors identify several critical gaps in the current state of the art:

  • Multilingual and Low-Resource Gaps: The vast majority of tools, datasets, and benchmarks are optimized for English. Robust anonymization for low-resource languages remains a significant hurdle.
  • Static vs. Dynamic Anonymization: Most current systems apply a blanket, static layer of anonymization. Future research must focus on adaptive systems that adjust the level of privacy based on context, data sensitivity, and user permissions.
  • Explainability: As anonymization models grow more complex, understanding why a model chose to redact or keep a specific piece of text is vital for trust and debugging.
  • Multimodal Expansion: Principles established for text need to be effectively translated to handle multimodal data (text mixed with images, audio, or code).

Summary updated

Abstract

The proliferation of textual data containing sensitive personal information across various domains requires robust anonymization techniques to protect privacy and comply with regulations, while preserving data usability for diverse and crucial downstream tasks. This survey provides a comprehensive overview of current trends and recent advances in text anonymization techniques. We begin by discussing foundational approaches, primarily centered on Named Entity Recognition, before examining the transformative impact of Large Language Models, detailing their dual role as sophisticated anonymizers and potent de-anonymization threats. The survey further explores domain-specific challenges and tailored solutions in critical sectors such as healthcare, law, finance, and education. We investigate advanced methodologies incorporating formal privacy models and risk-aware frameworks, and address the specialized subfield of authorship anonymization. Additionally, we review evaluation frameworks, comprehensive metrics, benchmarks, and practical toolkits for real-world deployment of anonymization solutions. This review consolidates current knowledge, identifies emerging trends and persistent challenges, including the evolving privacy-utility trade-off, the need to address quasi-identifiers, and the implications of LLM capabilities, and aims to guide future research directions for both academics and practitioners in this field.
BibTeX
@misc{deuer2025survey,
  author = {Tobias Deußer and Lorenz Sparrenberg and Armin Berger and Max Hahnbück and Christian Bauckhage and Rafet Sifa},
  title = {{A Survey on Current Trends and Recent Advances in Text Anonymization}},
  year = {2025},
  eprint = {2508.21587},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.21587},
}
Figure