mlsanityData to 14 Oct 2026

Differentially Private Learning Needs Better Model Initialization and Self-Distillation

arXiv 2410.17566 ↗First posted 23 Oct 2024cs.LG cs.AI cs.CLcited by 3

On the map: Knowledge distillation → Federated, robust and private distillation

Abstract (arXiv)

Differentially private SGD (DPSGD) enables privacy-preserving training of language models, but often reduces utility, diversity, and linguistic quality. We introduce DPRefine, a three-phase method that initializes a model using data synthesis from a small pre-trained LM with rigorous filtering, applies DP finetuning on private data, and performs self-distillation to refine outputs. This approach significantly outperforms vanilla DPSGD, with AlpacaEval preferring DPRefine's generations in 78.4% of cases across all datasets. Our analysis reveals that DPRefine reduces linguistic errors in generated text by 84.0%, mitigating grammar and spelling errors, commonly associated with DPSGD. It also reduces inconsistencies of non-private models, such as hallucinated details and misattributed quotes. We find that small models like GPT-2 can be effective for initialization and distillation, highlighting their potential in enabling scalable and efficient deployment of privacy-preserving language.

Key facts. Cited by 3 papers in the map (data through 14 Oct 2026). Placed in Federated, robust and private distillation, one of 12 directions in Knowledge distillation. Builds on 30 of its 56 references that are in the map. First posted on arXiv 23 Oct 2024.

Builds on

30 of 56 references.

Closest ancestors · 1

Same direction, newest first

  1. 2024Differentially Private Knowledge Distillation via Synthetic Text Generation1 Mar

Other references · 29

  1. 2019BERTScore: Evaluating Text Generation with BERTcited by 5,512
  2. 2018GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understandingcited by 4,899
  3. 2016Deep Learning with Differential Privacycited by 2,728
  4. 2018Don't Give Me the Details, Just the Summary! Topic-Aware Convolutional Neural Networks for Extreme Summarizationcited by 1,317
  5. 2019TinyBERT: Distilling BERT for Natural Language Understandingcited by 1,167
  6. 2022BioGPT: Generative Pre-trained Transformer for Biomedical Text Generation and Miningcited by 538
  7. 2019Patient Knowledge Distillation for BERT Model Compressioncited by 508
  8. 2023Benchmarking Large Language Models for News Summarizationcited by 445
  9. 2021Large Language Models Can Be Strong Differentially Private Learnerscited by 412
  10. 2021Differentially Private Fine-tuning of Language Modelscited by 400
  11. 2019Differential Privacy Has Disparate Impact on Model Accuracycited by 316
  12. 2022Data Contamination: From Memorization to Exploitationcited by 236
17 more
  1. 2020Differentially Private Learning Needs Better Features (or Much More Data)cited by 235
  2. 2022WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creationcited by 202
  3. 2023What's In My Big Data?cited by 129
  4. 2024Best Practices and Lessons Learned on Synthetic Datacited by 106
  5. 2022Position: Considerations for Differentially Private Learning with Large-Scale Public Pretrainingcited by 88
  6. 2024Unfamiliar Finetuning Examples Control How Language Models Hallucinatecited by 87
  7. 2023Towards Mitigating Hallucination in Large Language Models via Self-Reflectioncited by 81
  8. 2024Differentially Private Synthetic Data via Foundation Model APIs 2: Textcited by 76
  9. 2024TofuEval: Evaluating Hallucinations of LLMs on Topic-Focused Dialogue Summarizationcited by 66
  10. 2024Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Samplingcited by 61
  11. 2022When Does Differentially Private Learning Not Suffer in High Dimensions?cited by 48
  12. 2020Differentially Private Language Models Benefit from Public Pre-trainingcited by 47
  13. 2016Revisiting Summarization Evaluation for Scientific Articlescited by 31
  14. 2021Privacy Regularization: Joint Privacy-Utility Optimization in Language Modelscited by 30
  15. 2020Towards Differentially Private Text Representationscited by 29
  16. 2023Advancing Differential Privacy: Where We Are Now and Future Directions for Real-World Deploymentcited by 27
  17. 2022Privacy-Preserving Domain Adaptation of Semantic Parserscited by 12

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 1; 2026: 2.

3 papers.

  1. 2025Differentially-private text generation degrades output language qualitycited by 1
  2. 2026DP-RFT: Learning to Generate Synthetic Text via Differentially Private Reinforcement Fine-Tuningcited by 0
  3. 2026The Privacy-Hallucination Tradeoff in Differentially Private Language Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026DP-OPD: Differentially Private On-Policy Distillation for Language Models
  2. 2023DP-Forward: Fine-tuning and Inference on Language Models with Differential Privacy in Forward Pass
  3. 2024SafeSynthDP: Leveraging Large Language Models for Privacy-Preserving Synthetic Data Generation Using Differential Privacy
  4. 2025DP-Fusion: Token-Level Differentially Private Inference for Large Language Models
  5. 2025Can Differentially Private Fine-tuning LLMs Protect Against Privacy Attacks?
  6. 2021An Efficient DP-SGD Mechanism for Large Scale NLP Models
  7. 2025Dual-Priv Pruning : Efficient Differential Private Fine-Tuning in Multimodal Large Language Models
  8. 2025Public Data Assisted Differentially Private In-Context Learning