Privacy-preserving data sharing via probabilistic modelling.

Joonas Jälkö, Eemil Lagerspetz, Jari Haukka, Sasu Tarkoma, Samuel Kaski, Antti HonkelaarXiv 2019 · DOI · OpenReview

Abstract

Differential privacy allows quantifying privacy loss resulting from accessing sensitive personal data. Repeated accesses to underlying data incur increasing loss. Releasing data as privacy-preserving synthetic data would avoid this limitation, but would leave open the problem of designing what kind of synthetic data. We propose formulating the problem of private data release through probabilistic modelling. This approach transforms the problem of designing the synthetic data into choosing a model for the data, allowing also including prior knowledge, which improves the quality of the synthetic data. We demonstrate empirically, in an epidemiological study, that statistical discoveries can be reliably reproduced from the synthetic data. We expect the method to have broad use in creating high-quality anonymized data twins of key data sets for research.

Introduction

Open data sharing is a cornerstone of scientific progress but is often hindered by the sensitive nature of personal data, such as clinical or genomic records. Differentially private (DP) synthetic data offers a mathematical guarantee of privacy while aiming to preserve the statistical utility of the original dataset. However, standard black-box generative models often struggle with small datasets or weak signals, as they exhaust their limited "privacy budget" on learning trivial or obvious data structures.

The Probabilistic Modelling Framework

This paper proposes a "Bayesian DP data release" framework that treats synthetic data generation as a probabilistic modelling task. Instead of treating the generator as a black box, this approach allows researchers to incorporate domain knowledge through probabilistic modelling before the privacy-preserving training begins.

The proposed workflow for private data release using probabilistic modelling, incorporating domain-specific models and priors.

By defining a model $p(\mathbf{X}, \boldsymbol{\theta})$ that encodes known constraints—such as structural zeros, known stratifications, or conditional dependencies—the private learning process can focus its limited information-extraction capacity on the unknown parameters that drive scientific discovery. The workflow involves three main steps:

  1. Model Definition: A domain expert defines a probabilistic model (e.g., a Mixture Model) with informative priors.
  2. Private Learning: The posterior distributions of the model parameters are learned using Differentially Private Variational Inference (DPVI), which utilizes DP-SGD to bound privacy loss.
  3. Generation: Synthetic data is sampled from the resulting posterior predictive distribution (PPD).

Experimental Evaluation

The framework was evaluated on several real-world datasets, including the UCI Adult benchmark and a high-dimensional mobile app dataset (Carat). Most notably, it was tested on an epidemiological study of Alcohol-Related Deaths (ARD) involving over 400,000 records. The goal was to determine if the core medical discovery—the association between diabetes medication and alcohol-related mortality—could be reliably reproduced using synthetic data.

Reproducibility of discoveries in the ARD study for males (a) and females (b) across different privacy levels ($\epsilon$) and stratification strategies.

Key results from the ARD study showed that encoding prior knowledge (stratification) significantly improved reproducibility. For males, under a strict privacy budget ($\epsilon = 1$), the "Stratified" model increased the rate of reproduced discoveries by nearly 40% compared to unstratified approaches. However, the study also highlighted that rare events or weak signals (as seen in the female cohort, where the original signal was weaker) remain challenging for DP mechanisms, as noise can easily overwhelm the signal in small sub-populations.

Key Insights

  • Priors Optimize Privacy Budgets: Hardcoding known constraints prevents the model from wasting the privacy budget on obvious facts, leaving more room to capture subtle, valuable statistical signals.
  • Synthetic Data vs. Tailored Queries: Traditional DP requires splitting a privacy budget for every query. Releasing a single high-quality synthetic dataset can be more efficient for multi-purpose downstream analysis, as the privacy cost is paid only once upfront.
  • Scalability: While Bayesian networks are effective for low-dimensional data, the proposed mixture model approach scales significantly better to high-dimensional datasets while maintaining lower error rates.

Conclusion

By transforming synthetic data generation into a probabilistic modelling task, this approach enables the creation of high-utility "data twins." This method is particularly promising for research datasets where domain expertise can be leveraged to maximize the utility of sensitive data under rigorous privacy guarantees. Future work may further optimize hyperparameter selection under DP constraints to enhance the robustness of the generative process.

Summary updated

Abstract

Differential privacy allows quantifying privacy loss resulting from accessing sensitive personal data. Repeated accesses to underlying data incur increasing loss. Releasing data as privacy-preserving synthetic data would avoid this limitation, but would leave open the problem of designing what kind of synthetic data. We propose formulating the problem of private data release through probabilistic modelling. This approach transforms the problem of designing the synthetic data into choosing a model for the data, allowing also including prior knowledge, which improves the quality of the synthetic data. We demonstrate empirically, in an epidemiological study, that statistical discoveries can be reliably reproduced from the synthetic data. We expect the method to have broad use in creating high-quality anonymized data twins of key data sets for research.
BibTeX
@misc{jalko2019privacy,
  author = {Joonas Jälkö and Eemil Lagerspetz and Jari Haukka and Sasu Tarkoma and Samuel Kaski and Antti Honkela},
  title = {{Privacy-preserving data sharing via probabilistic modelling.}},
  year = {2019},
  eprint = {1912.04439},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1912.04439},
}
Figure