mlsanityData to 7 Oct 2026

Privacy Regularization: Joint Privacy-Utility Optimization in Language Models

arXiv 2103.07567 ↗First posted 12 Mar 2021cs.LG cs.CL cs.CRcited by 29

Not on the map. Why?

Abstract (arXiv)

Neural language models are known to have a high capacity for memorization of training samples. This may have serious privacy implications when training models on user content such as email correspondence. Differential privacy (DP), a popular choice to train models with privacy guarantees, comes with significant costs in terms of utility degradation and disparate impact on subgroups of users. In this work, we introduce two privacy-preserving regularization methods for training language models that enable joint optimization of utility and privacy through (1) the use of a discriminator and (2) the inclusion of a triplet-loss term. We compare our methods with DP through extensive evaluation. We show the advantages of our regularizers with favorable utility-privacy trade-off, faster training with the ability to tap into existing optimization approaches, and ensuring uniform treatment of under-represented subgroups.

Key facts. Cited by 29 papers in the map (data through 7 Oct 2026). Builds on 16 of its 30 references that are in the map. First posted on arXiv 12 Mar 2021.

Builds on

16 of 30 references.

References · 16

  1. 2016Deep Learning with Differential Privacycited by 2,682
  2. 2016Membership Inference Attacks against Machine Learning Modelscited by 2,221
  3. 2020Extracting Training Data from Large Language Modelscited by 2,144
  4. 2018The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networkscited by 847
  5. 2019Differential Privacy Has Disparate Impact on Model Accuracycited by 316
  6. 2020Differentially Private Learning Needs Better Features (or Much More Data)cited by 233
  7. 2019Deep Learning with Gaussian Differential Privacycited by 112
  8. 2019Gmail Smart Compose: Real-Time Assisted Writingcited by 97
  9. 2018A General Approach to Adding Differential Privacy to Iterative Training Procedurescited by 84
  10. 2018Towards Robust and Privacy-preserving Text Representationscited by 75
  11. 2020Privacy in Deep Learning: A Surveycited by 70
  12. 2020Training Production Language Models without Memorizing User Datacited by 68
4 more
  1. 2020Neither Private Nor Fair: Impact of Data Imbalance on Utility and Fairness in Differential Privacycited by 61
  2. 2019Fair Decision Making using Privacy-Protected Datacited by 58
  3. 2020Enabling Fast Differentially Private SGD via Just-in-Time Compilation and Vectorizationcited by 58
  4. 2018Privacy-preserving Neural Representations of Textcited by 54

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2021: 4; 2022: 8; 2023: 6; 2024: 3; 2025: 5; 2026: 3.

29 papers.

Sort:

By area:

  1. 2021Large Language Models Can Be Strong Differentially Private Learnerscited by 405
  2. 2022Deduplicating Training Data Mitigates Privacy Risks in Language Modelscited by 309
  3. 2023Membership Inference Attacks against Language Models via Neighbourhood Comparisoncited by 249
  4. 2022Quantifying Privacy Risks of Masked Language Models Using Membership Inference Attackscited by 168
  5. 2023Practical Membership Inference Attacks against Fine-tuned Large Language Models via Self-prompt Calibrationcited by 90
  6. 2022Do Language Models Plagiarize?cited by 72
  7. 2022Mix and Match: Learning-free Controllable Text Generation using Energy Language Modelscited by 66
  8. 2022A Customized Text Sanitization Mechanism with Differential Privacycited by 51
  9. 2023Training Data Extraction From Pre-trained Language Models: A Surveycited by 43
  10. 2021Membership Inference on Word Embedding and Beyondcited by 34
  11. 2022Submix: Practical Private Prediction for Large-Scale Language Modelscited by 31
  12. 2024PrE-Text: Training Language Models on Private Federated Data in the Age of LLMscited by 29
17 more
  1. 2021Benchmarking Differential Privacy and Federated Learning for BERT Modelscited by 28
  2. 2024Navigating LLM Ethics: Advancements, Challenges, and Future Directionscited by 28
  3. 2023Are Chatbots Ready for Privacy-Sensitive Applications? An Investigation into Input Regurgitation and Prompt-Induced Sanitizationcited by 17
  4. 2021Style Pooling: Automatic Text Style Obfuscation for Improved Classification Fairnesscited by 12
  5. 2022You Don't Know My Favorite Color: Preventing Dialogue Representations from Revealing Speakers' Private Personascited by 11
  6. 2022Privacy Leakage in Text Classification: A Data Extraction Approachcited by 9
  7. 2025Detecting Data Contamination in LLMs via In-Context Learningcited by 7
  8. 2023Deconstructing Classifiers: Towards A Data Reconstruction Attack Against Text Classification Modelscited by 6
  9. 2024Differentially Private Learning Needs Better Model Initialization and Self-Distillationcited by 3
  10. 2025Learning Obfuscations Of LLM Embedding Sequences: Stained Glass Transformcited by 2
  11. 2023Using Property Elicitation to Understand the Impacts of Fairness Regularizerscited by 0
  12. 2025Fragments to Facts: Partial-Information Fragment Inference from LLMscited by 0
  13. 2025UNO: Unlearning via Orthogonalization in Generative modelscited by 0
  14. 2025RL-Finetuned LLMs for Privacy-Preserving Synthetic Rewritingcited by 0
  15. 2026Chain-of-Authorization: Embedding authorization into large language modelscited by 0
  16. 2026FISGuard: Defending Against Membership Inference via Fixed Input Subspacescited by 0
  17. 2026The Privacy-Hallucination Tradeoff in Differentially Private Language Modelscited by 0