Exploring The Empirical Privacy Landscape of Synthetic Data
Summary
该论文对合成数据上的隐私风险进行了实证研究，比较了数据级与训练级干预手段在能力与隐私之间的权衡，发现隐私感知重写相比 DP-SGD 或直接微调能提供更优的能力-隐私平衡。
View Cached Full Text
Cached at: 10/22/26, 07:36 PM
Paper page - Exploring The Empirical Privacy Landscape of Synthetic Data
Source: https://huggingface.co/papers/2610.22659
Abstract
Privacy-aware rewriting of synthetic medical records improves privacy protection with less capability loss than direct fine-tuning or DP-SGD, offering a favorable trade-off for sensitive data.
Methods for training language models on sensitive data differ in how they balance capability gains against the risk of leaking private information. In this paper, we present a controlled comparison of data-level and training-level interventions for navigating the capability-privacy tradeoff onsynthetic medical recordswithcanariesto probe for leakage. Direct finetuning performs best on capability metrics but leads to the most privacy leakage, whileDP-SGDreduces leakage at substantial cost to capabilities. Generating synthetic data from these records offers a more favorable balance, particularly when guided by auniqueness-aware instructionto replace private details. Relative to naive paraphrasing,privacy-aware rewritingimproves privacy protection across a variety of different rewriting models. We also find that more capable rewriting models generally reduce the risks of downstream privacy leakage, althoughheld-out fitdoes not improve monotonically and increasedtest-time computeyields inconsistent gains. Overall, these results highlightprivacy-aware rewritingas a promising direction for achieving a more favorable trade-off between capabilities and empirical privacy compared to other training-level methods, and our evaluation framework supports comparisons of future models and methods.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2610\.22659
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2610.22659 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2610.22659 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2610.22659 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Revisiting the Provable-Auditable Privacy Gap of DP-SGD
The paper proposes a lightweight defense framework to enhance the empirical privacy of DP-SGD without theoretical privacy cost, validated through extensive audits across models, datasets, and threat models.
Gaussian Approximation for Composition and Subsampling of Differentially Private Algorithms
该论文证明了差分隐私算法在组合与子采样下的隐私损失分布可通过截断后的高斯近似来刻画，从而为DP-SGD等场景给出比Rényi分析更紧的闭式隐私界，尤其在采样率1/t时显著降低所需噪声规模σ的要求。
Evaluating LLM Simulators as Differentially Private Data Generators
This paper evaluates LLM-based simulators as generators of differentially private synthetic data, using PersonaLedger to assess whether LLMs can faithfully reproduce statistical distributions from DP-protected personas. While achieving promising fraud detection utility (AUC 0.70 at ε=1), the study identifies significant distribution drift caused by systematic LLM biases that override input statistics.
SHIP: Structured and Hierarchical Insight with Privacy
该论文提出SHIP（Structured and Hierarchical Insight with Privacy），将差分隐私洞察形式化为可复用的查询回答表示，通过自动构建层级混合分类法并发布DP合成表，实现隐私保护下的文本数据分析，性能优于现有私有洞察基线。
Provable Robustness against Backdoor Attacks via the Primal-Dual Perspective on Differential Privacy
This paper introduces a framework that connects randomized smoothing to differential privacy through privacy profiles, enabling tight provable robustness guarantees against backdoor attacks that jointly affect training and inference. The approach is instantiated for DP-SGD and Deep Partition Aggregation with experiments on MNIST and CIFAR-10.