mlsanityData to 14 Oct 2026

QuanText: Protecting Dataset-Level Secrets in Textual Data Sharing

arXiv 2609.17995 ↗First posted 16 Sep 2026cs.LG cs.CRcited by 0

Not on the map. Why?

Abstract (arXiv)

Natural-language datasets support many downstream applications and research studies, but releasing text can reveal sensitive global properties of the underlying data source, such as the proportion of records associated with a particular gender, diagnosis, or political stance. Existing work has largely focused on property inference attacks that recover such global properties, while defenses for protecting these dataset-level secrets remain limited. Differential privacy, although effective for protecting individual records, provides only weak protection for aggregate properties. We propose Randomized Quantization for Text (QuanText), a training-free and large-language-model-agnostic data release mechanism that protects global secrets in textual datasets while preserving data utility. Given a dataset-level secret, such as the proportion of records with a particular diagnosis, and attributes whose utility should be preserved, such as topic and sentiment, QuanText perturbs both the secret distribution and the distributions of correlated attributes. It does so by constructing candidate release distributions over secret and non-secret attributes, randomly selecting a candidate sufficiently close to the private empirical distribution, and rewriting each private text sample to match the selected distribution using attribute-related snippets from the original text. QuanText is inspired by the Statistic Maximal Leakage (SML) framework, which bounds leakage about a secret function of a data distribution. Under idealized conditions, we show that QuanText satisfies an SML guarantee. Since these conditions may not hold exactly in practice, we also evaluate QuanText empirically on real-world datasets. Our results show that QuanText achieves a better empirical privacy-utility trade-off than competing data generation baselines.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 35 of its 52 references that are in the map. First posted on arXiv 16 Sep 2026.

Builds on

35 of 52 references.

References · 35

  1. 2016Deep Learning with Differential Privacycited by 2,728
  2. 2020Extracting Training Data from Large Language Modelscited by 2,189
  3. 2018Datasheets for Datasetscited by 1,386
  4. 2023DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Modelscited by 518
  5. 2021Differentially Private Fine-tuning of Language Modelscited by 400
  6. 2023ChatDoctor: A Medical Chat Model Fine-Tuned on a Large Language Model Meta-AI (LLaMA) Using Medical Domain Knowledgecited by 325
  7. 2023Analyzing Leakage of Personally Identifiable Information in Language Modelscited by 285
  8. 2017Style Transfer from Non-Parallel Text by Cross-Alignmentcited by 248
  9. 2017Style Transfer in Text: Exploration and Evaluationcited by 146
  10. 2013Hacking Smart Machines with Smarter Ones: How to Extract Meaningful Data from Machine Learning Classifierscited by 130
  11. 2018Disentangled Representation Learning for Non-Parallel Text Style Transfercited by 127
  12. 2022Synthetic Text Generation with Differential Privacy: A Simple and Practical Recipecited by 108
23 more
  1. 2024Differentially Private Synthetic Data via Foundation Model APIs 2: Textcited by 76
  2. 2023Differentially Private Synthetic Data via Foundation Model APIs 1: Imagescited by 65
  3. 2022Are Attribute Inference Attacks Just Imputation?cited by 49
  4. 2020Leakage of Dataset Properties in Multi-Party Machine Learningcited by 47
  5. 2021Formalizing and Estimating Distribution Inference Riskscited by 43
  6. 2022Are Your Sensitive Attributes Private? Novel Model Inversion Attribute Inference Attacks on Classification Modelscited by 43
  7. 2021Property Inference Attacks Against GANscited by 33
  8. 2021On the (In)Feasibility of Attribute Inference Attacks on Machine Learning Modelscited by 31
  9. 2022SNAP: Efficient Extraction of Private Properties with Poisoningcited by 31
  10. 2024PrE-Text: Training Language Models on Private Federated Data in the Age of LLMscited by 29
  11. 2022Distribution inference risks: Identifying and mitigating sources of leakagecited by 17
  12. 2022Inferring Sensitive Attributes from Model Explanationscited by 15
  13. 2022Dissecting Distribution Inferencecited by 15
  14. 2025Differentially Private Synthetic Data via APIs 3: Using Simulators Instead of Foundation Modelcited by 13
  15. 2025DPImageBench: A Unified Benchmark for Differentially Private Image Synthesiscited by 11
  16. 2021Correlation inference attacks against machine learning modelscited by 7
  17. 2025Struct-Bench: A Benchmark for Differentially Private Structured Text Generationcited by 7
  18. 2023Summary Statistic Privacy in Data Sharingcited by 6
  19. 2023PriSampler: Mitigating Property Inference of Diffusion Modelscited by 4
  20. 2025Can We Infer Confidential Properties of Training Data from LLMs?cited by 4
  21. 2026SynAE: A Framework for Measuring the Quality of Synthetic Data for Tool-Calling Agent Evaluationscited by 2
  22. 2026Differentially Private Synthetic Data via APIs 4: Tabular Datacited by 2
  23. 2024Guarding Multiple Secrets: Enhanced Summary Statistic Privacy for Data Sharingcited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Quantization Enables Private Dense Retrieval against Malicious Service Providers
  2. 2024PrivacyRestore: Privacy-Preserving Inference in Large Language Models via Privacy Removal and Restoration
  3. 2021ADePT: Auto-encoder based Differentially Private Text Transformation
  4. 2025Synthesizing Privacy-Preserving Text Data via Finetuning without Finetuning Billion-Scale LLMs
  5. 2023InferDPT: Privacy-Preserving Inference for Closed-box Large Language Model
  6. 2026TextCloak: Thwarting Unauthorized LLM Exploitation via RL-Driven Unlearnable Text
  7. 2025STAMP Your Content: Proving Dataset Membership via Watermarked Rephrasings
  8. 2025SynBench: A Benchmark for Differentially Private Text Generation