Pith. sign in

REVIEW 13 cited by

Differentially Private Fine-tuning of Language Models

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2110.06500 v2 pith:6IDSR27M submitted 2021-10-13 cs.LG cs.CLcs.CRstat.ML

classification cs.LGcs.CLcs.CRstat.ML
keywords privateprivacyfine-tuningmodelsaccuracyachievedifferentiallylanguage
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
abstract

We give simpler, sparser, and faster algorithms for differentially private fine-tuning of large-scale pre-trained language models, which achieve the state-of-the-art privacy versus utility tradeoffs on many standard NLP tasks. We propose a meta-framework for this problem, inspired by the recent success of highly parameter-efficient methods for fine-tuning. Our experiments show that differentially private adaptations of these approaches outperform previous private algorithms in three important dimensions: utility, privacy, and the computational and memory cost of private training. On many commonly studied datasets, the utility of private models approaches that of non-private models. For example, on the MNLI dataset we achieve an accuracy of $87.8\%$ using RoBERTa-Large and $83.5\%$ using RoBERTa-Base with a privacy budget of $\epsilon = 6.7$. In comparison, absent privacy constraints, RoBERTa-Large achieves an accuracy of $90.2\%$. Our findings are similar for natural language generation tasks. Privately fine-tuning with DART, GPT-2-Small, GPT-2-Medium, GPT-2-Large, and GPT-2-XL achieve BLEU scores of 38.5, 42.0, 43.1, and 43.8 respectively (privacy budget of $\epsilon = 6.8,\delta=$ 1e-5) whereas the non-private baseline is $48.1$. All our experiments suggest that larger models are better suited for private fine-tuning: while they are well known to achieve superior accuracy non-privately, we find that they also better maintain their accuracy when privacy is introduced.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 13 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. In-Context Probing for Membership Inference in Fine-Tuned Language Models

    cs.CR 2025-12 conditional novelty 6.0 of 10

    ICP-MIA infers membership in fine-tuned LLMs by measuring confidence improvement under in-context probes, beating prior black-box attacks at low false-positive rates.

  2. UniAud: A Unified Auditing Framework for High Auditing Power and Utility with One Training Run

    cs.CR 2025-07 conditional novelty 6.0 of 10

    UniAud uses synthetic uncorrelated canaries and self-comparison inference to reach near-optimal empirical epsilon lower bounds in one black-box DP audit run, while UniAud++ improves the utility-auditing trade-off via ...

  3. Square$\chi$PO: Differentially Private and Robust $\chi^2$-Preference Optimization in Offline Direct Alignment

    cs.LG 2025-05 conditional novelty 6.0 of 10

    SquareχPO, a square-loss variant of χPO, achieves optimal 1/sqrt(n) suboptimality under label privacy and Huber corruption for offline direct alignment with general function classes.

  4. GROM: Gradient-Free Rapid One-Shot Machine Unlearning

    cs.LG 2026-08 conditional novelty 5.0 of 10

    A single closed-form ridge update to selected MLP layers removes targeted knowledge from LLMs in seconds, with state-of-the-art forgetting-utility trade-offs and quantization robustness.

  5. Membership Inference Attacks on Tokenizers of Large Language Models

    cs.CR 2025-10 conditional novelty 5.0 of 10

    Tokenizers leak dataset membership: using distinctive tokens that appear in a dataset's text, an attacker can detect with AUC up to 0.77 whether that dataset was part of tokenizer training.

  6. Re-examining Low Rank adaptation for private LLM fine-tuning

    cs.LG 2025-10 conditional novelty 5.0 of 10

    Denoising DP-SGD gradients with random-matrix-theory singular-value shrinkage speeds up private RoBERTa fine-tuning on GLUE, cutting steps to target accuracy by 20–100%.

  7. Differentially-private text generation degrades output language quality

    cs.CL 2025-09 conditional novelty 5.0 of 10

    DP fine-tuning systematically degrades LLM output length, grammatical correctness, and lexical diversity, and this degradation grows as the privacy budget shrinks.

  8. Balancing Utility and Privacy: Dynamically Private SGD with Random Projection

    cs.LG 2025-09 reject novelty 5.0 of 10

    D2P2-SGD combines time-decreasing privacy noise with random projection to improve the accuracy of differentially private SGD, with convergence rates matching ordinary SGD.

  9. On the Performance of Differentially Private Optimization with Heavy-Tail Class Imbalance

    cs.LG 2025-07 conditional novelty 5.0 of 10

    Under heavy-tail class imbalance, subtracting the DP noise variance from Adam's second moment (DP-AdamBC) substantially improves learning of rare classes compared with DP gradient descent.

  10. Dual-Priv Pruning : Efficient Differential Private Fine-Tuning in Multimodal Large Language Models

    cs.CR 2025-06 conditional novelty 5.0 of 10

    A framework for DP fine-tuning of MLLMs that prunes visual tokens before training and selectively applies noisy gradient updates to blocks with the largest norms, reporting modest utility and memory gains over DP-SGD.

  11. Context Reasoner: Incentivizing Reasoning Capability for Contextualized Privacy and Safety Compliance via Reinforcement Learning

    cs.CL 2025-05 conditional novelty 5.0 of 10

    A Context Reasoner pipeline that cold-starts LLMs on distilled legal reasoning and applies PPO with a rule-based compliance reward improves performance on CI-based legal compliance benchmarks and transfers to general ...

  12. Public Data Assisted Differentially Private In-Context Learning

    cs.AI 2025-09 conditional novelty 4.0 of 10

    A private ICL algorithm that aggregates LLM responses with DPM clustering and uses public data representatives achieves near-non-private utility at epsilon=1.

  13. How to DP-fy Your Data: A Practical Guide to Generating Synthetic Data With Differential Privacy

    cs.CR 2025-12 conditional novelty 2.0 of 10

    A practical, extremely thorough survey of differentially private synthetic data generation: methods, privacy units, evaluation metrics, and end-to-end system components across four data modalities.

Pith tools