Privacy-preserving data sharing via probabilistic modeling

Joonas Jälkö, Eemil Lagerspetz, Jari Haukka, Sasu Tarkoma, Antti Honkela, Samuel KaskiPatterns 2021 · OpenReview

TL;DR

This paper introduces a novel approach to privacy-preserving data sharing by framing synthetic data generation as a probabilistic modeling task. By incorporating prior domain knowledge into generative models, they significantly improve the quality and utility of synthetic data while maintaining strong privacy guarantees, addressing a key limitation of existing methods.

What Problem Does This Paper Address?

The increasing demand for open data in scientific research is often at odds with the need to protect sensitive personal information. While releasing synthetic data is a promising solution, existing methods struggle to balance privacy with data utility. The core challenge is that privacy mechanisms, like differential privacy (DP), often introduce noise that degrades the quality of synthetic data, making it less useful for downstream analysis. Furthermore, these methods typically learn models from scratch, failing to leverage valuable existing knowledge about the data generation process, which could otherwise improve efficiency and accuracy. This paper aims to bridge this gap by enabling the incorporation of prior knowledge into privacy-preserving synthetic data generation.

Methods

The authors propose a framework they call "Bayesian DP data release." This approach reframes synthetic data generation as a modeling problem, allowing the integration of prior knowledge before the privacy-preserving learning process begins. Instead of learning a generative model from scratch, they augment the standard workflow by allowing users to encode their understanding of the data's structure and generation process into a probabilistic model.

Key aspects of their method include:

  • Probabilistic Modeling: They leverage probabilistic models, which provide a natural language for expressing prior knowledge. This allows for the integration of domain expertise, such as structural zeros (e.g., living subjects cannot have a cause of death) or relationships between variables.
  • Generative Models: The core of their approach involves learning a generative model from the sensitive data under DP guarantees and then sampling synthetic data from this trained model.
  • Posterior Predictive Distribution (PPD): They use the PPD to generate synthetic data, which represents the probability of observing new samples conditioned on the data observed so far. This is achieved by first sampling from the posterior distribution of the model parameters and then sampling new data points from the probabilistic model.
  • Differential Privacy (DP): The entire learning process is conducted under DP to ensure privacy guarantees. They use approximate DP, with specific epsilon (ε) and delta (δ) values chosen based on the dataset and desired privacy level.

Datasets

The authors evaluated their approach on three datasets:

  • ARD (Alcohol-Related Deaths) Study: An epidemiological dataset used to study the association between diabetes and alcohol-related deaths. This dataset was used extensively to demonstrate the impact of incorporating prior knowledge.
  • Carat Dataset: A mobile phone app dataset used for understanding battery usage. This dataset was used to evaluate performance against tailored mechanisms and compare different modeling approaches.
  • Adult Dataset (UCI Machine Learning Repository): A well-known dataset used for predicting income classes. This dataset was also used for evaluating performance against tailored mechanisms and comparing modeling approaches.

No new benchmarks were introduced; rather, existing datasets were used to demonstrate the effectiveness of their proposed method.

Key Insights and Ideas

  • Prior Knowledge is Crucial: The most significant insight is that encoding prior knowledge into the generative model dramatically improves the quality and utility of synthetic data, often more so than simply increasing the privacy budget (ε).
  • Modeling as a Solution: Viewing synthetic data generation as a modeling task, rather than just a direct data-driven learning problem, unlocks the potential to incorporate domain expertise effectively.
  • Balancing Privacy and Utility: The paper demonstrates that by intelligently incorporating prior knowledge, it's possible to achieve strong privacy guarantees while mitigating the quality degradation typically associated with DP.
  • Weak Signals are Challenging: Even with their improved method, capturing weak statistical signals in the data can be difficult, especially under strict privacy constraints. This highlights a fundamental trade-off in DP.
  • Synthetic Data for Multiple Uses: Releasing synthetic data offers a significant advantage over traditional DP query-based methods when data needs to be accessed multiple times, as it avoids the depletion of a privacy budget.

Key Results

  • Improved Reproducibility: In the ARD study, encoding prior knowledge (gender-specific models and outcome-based stratification) significantly increased the probability of reproducing statistical discoveries from synthetic data compared to models without such knowledge. For males, reproducibility increased by nearly 40% with strict privacy (ε=1).
  • Performance Comparison: Mixture models generally outperformed Bayes networks in terms of accuracy for higher-dimensional data (Carat study) and better preserved regression coefficients (ARD study). However, Bayes networks showed better classification accuracy on the Adult dataset.
  • Efficiency: Mixture models were found to be computationally more efficient than Bayes networks, especially as dimensionality increased.
  • Utility vs. Privacy: The study showed that for females in the ARD study, looser privacy guarantees (higher ε) were needed to reproduce discoveries with high probability, indicating a trade-off between privacy and utility, especially for smaller subgroups or weaker signals.
  • Comparison to Tailored Mechanisms: For multiple future queries, their data-sharing methods (mixture model and Bayes network) outperformed traditional tailored DP mechanisms (e.g., Gaussian mechanism) in terms of accuracy.

Limitations and Future Work

  • Weak Signals: The paper acknowledges that capturing weak statistical signals remains challenging, particularly for smaller subgroups or when strict privacy is enforced.
  • Computational Cost: While mixture models are more efficient than Bayes networks, learning complex probabilistic models under DP can still be computationally intensive.
  • Hyperparameter Tuning: Optimizing hyperparameters for DPVI (Differentially Private Variational Inference) can be complex and may incur additional privacy costs.
  • Future Work: The authors suggest that their approach has broad applicability for creating high-quality anonymized data twins of key datasets. Further research could explore more sophisticated ways to encode prior knowledge and investigate the theoretical underpinnings of how prior knowledge interacts with DP. They also highlight the need for cryptographically secure random number generators and careful consideration of floating-point arithmetic limitations in DP implementations.

Summary updated

Abstract

Generating synthetic data using privacy-preserving models is a promising method for sharing sensitive data. This paper proposes to view synthetic data generation through probabilistic modeling, which allows the improvement of data generation by incorporating prior knowledge into the generative model. The proposed approach allows us to counteract reduction in quality, which results from the obfuscation required for privacy, and as a result produces high-quality synthetic data with strong privacy guarantees.
BibTeX
@inproceedings{jalko2021privacy,
  author = {Joonas Jälkö and Eemil Lagerspetz and Jari Haukka and Sasu Tarkoma and Antti Honkela and Samuel Kaski},
  title = {{Privacy-preserving data sharing via probabilistic modeling}},
  booktitle = {Patterns},
  year = {2021},
  url = {https://openreview.net/forum?id=9jGIfkkQldC},
}
Figure