mlsanityData to 14 Oct 2026

PACZero: PAC-Private Fine-Tuning of Language Models via Sign Quantization

arXiv 2605.06505 ↗First posted 7 May 2026cs.LG cs.AI cs.CRcited by 1

Not on the map. Why?

Abstract (arXiv)

We introduce PACZero, a family of PAC-private zeroth-order mechanisms for fine-tuning large language models that delivers usable utility at I(S^*; Y1:T)=0. This privacy regime bounds the membership-inference attack (MIA) posterior success rate at the prior, an MIA-resistance level the DP framework matches only at varepsilon=0 and infinite noise. All DP-ZO comparisons below are matched at the MIA posterior level. The key insight is that PAC Privacy charges mutual information only when the release depends on which candidate subset is the secret. Sign-quantizing subset-aggregated zeroth-order gradients creates frequent unanimity, steps at which every candidate subset agrees on the update direction; at these steps the released sign costs zero conditional mutual information. We propose two variants that span the privacy-utility trade-off: PACZero-MI (budgeted MI via exact calibration on the binary release) and PACZero-ZPL (I=0 via a uniform coin flip on disagreement steps). We evaluate on SST-2 and SQuAD with OPT-1.3B and OPT-6.7B in both LoRA and full-parameter tracks. On SST-2 OPT-1.3B full fine-tuning at I=0, PACZero-ZPL reaches 88.99±0.91, within 2.1pp of the non-private MeZO baseline (91.1 FT). No prior method produces usable utility in the high-privacy regime varepsilon<1, and PACZero-ZPL obtains competitive SST-2 accuracy and nontrivial SQuAD F1 across OPT-1.3B and OPT-6.7B at I=0.

Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Builds on 34 of its 49 references that are in the map. First posted on arXiv 7 May 2026.

Builds on

34 of 49 references.

References · 34

  1. 2021LoRA: Low-Rank Adaptation of Large Language Modelscited by 14,380
  2. 2016SQuAD: 100,000+ Questions for Machine Comprehension of Textcited by 4,148
  3. 2022OPT: Open Pre-trained Transformer Language Modelscited by 3,512
  4. 2016Deep Learning with Differential Privacycited by 2,728
  5. 2016Membership Inference Attacks against Machine Learning Modelscited by 2,281
  6. 2020Extracting Training Data from Large Language Modelscited by 2,189
  7. 2021Membership Inference Attacks From First Principlescited by 878
  8. 2017Renyi Differential Privacycited by 650
  9. 2018signSGD: Compressed Optimisation for Non-Convex Problemscited by 571
  10. 2021Large Language Models Can Be Strong Differentially Private Learnerscited by 412
  11. 2021Differentially Private Fine-tuning of Language Modelscited by 400
  12. 2021Opacus: User-Friendly Differential Privacy Library in PyTorchcited by 383
22 more
  1. 2023Fine-Tuning Language Models with Just Forward Passescited by 345
  2. 2019Differential Privacy Has Disparate Impact on Model Accuracycited by 316
  3. 2013The Composition Theorem for Differential Privacycited by 297
  4. 2022Unlocking High-Accuracy Differentially Private Image Classification through Scalecited by 224
  5. 2022Quantifying Privacy Risks of Masked Language Models Using Membership Inference Attackscited by 175
  6. 2023How to DP-fy ML: A Practical Guide to Machine Learning with Differential Privacycited by 153
  7. 2022Automatic Clipping: Differentially Private Deep Learning Made Easier and Strongercited by 88
  8. 2022Differentially Private Optimization on Large Model at Small Costcited by 65
  9. 2020Individual Privacy Accounting via a Renyi Filtercited by 59
  10. 2022Exploring the Limits of Differentially Private Deep Learning with Group-wise Clippingcited by 52
  11. 2023Differentially Private Low-Rank Adaptation of Large Language Model Using Federated Learningcited by 41
  12. 2024Private Fine-tuning of Large Language Models with Zeroth-order Optimizationcited by 41
  13. 2022Individual Privacy Accounting for Differentially Private Stochastic Gradient Descentcited by 20
  14. 2022PAC Privacy: Automatic Privacy Measurement and Control of Data Processingcited by 19
  15. 2020Robustness Threats of Differential Privacycited by 15
  16. 2023Differentially Private SGD Without Clipping Bias: An Error-Feedback Approachcited by 14
  17. 2021Gradient Masking and the Underestimated Robustness Threats of Differential Privacy in Deep Learningcited by 13
  18. 2021Towards Understanding the Impact of Model Size on Differential Private Classificationcited by 10
  19. 2025An Improved Privacy and Utility Analysis of Differentially Private SGD with Bounded Domain and Smooth Lossescited by 4
  20. 2026Fundamental Limitations of Favorable Privacy-Utility Guarantees for DP-SGDcited by 4
  21. 2026SIMD-PAC-DB: Pretty Performant PAC Privacycited by 2
  22. 2026Private Prediction via PAC Privacycited by 1

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026Protecting Sensitive Data in Image Synthesis via PAC-Private Adaptation for Diffusion Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2024Efficient and Private: Memorisation under differentially private parameter-efficient fine-tuning in language models
  2. 2025Empirical Calibration and Metric Differential Privacy in Language Models
  3. 2024Privacy-Preserving Large Language Models: Mechanisms, Applications, and Future Directions
  4. 2025Dual-Priv Pruning : Efficient Differential Private Fine-Tuning in Multimodal Large Language Models
  5. 2024Privacy in Fine-tuning Large Language Models: Attacks, Defenses, and Future Directions
  6. 2025DP-Fusion: Token-Level Differentially Private Inference for Large Language Models
  7. 2025AlignDP: Hybrid Differential Privacy with Rarity-Aware Protection for LLMs
  8. 2023DP-Forward: Fine-tuning and Inference on Language Models with Differential Privacy in Forward Pass