Differential Privacy for Bayesian Inference through Posterior Sampling

Christos Dimitrakakis, Blaine Nelson, Zuhe Zhang, Aikaterini Mitrokotsa, Benjamin I. P. Rubinstein

TL;DR

This paper shows how to achieve differential privacy in Bayesian inference by carefully choosing the prior distribution. By sampling from a "private" posterior distribution, we can answer queries while protecting sensitive data, offering a trade-off between privacy and utility.

What Problem Does This Paper Address?

In many real-world scenarios, data is sensitive, and we want to extract useful information from it without revealing private details. Bayesian inference is a powerful framework for modeling uncertainty and making predictions. However, releasing information derived from a posterior distribution (which is based on the data) can inadvertently leak private information. The challenge is to develop Bayesian inference methods that are both useful and differentially private, meaning they provide strong privacy guarantees. This paper aims to bridge this gap by exploring how the choice of prior distribution can be leveraged to ensure privacy in Bayesian inference.

Methods

The core idea is to use posterior sampling as a mechanism for providing private responses to queries. Instead of releasing the exact posterior distribution, which could be sensitive, the authors propose drawing multiple samples from it. These samples then form a private representation of the posterior.

Key methodological contributions include:

  • Generalized Differential Privacy: The authors extend the definition of differential privacy to work with arbitrary data set metrics and outcome spaces, allowing for more flexible and realistic privacy analyses beyond standard tabular data.
  • Robustness and Privacy Link: They establish a connection between the robustness of the posterior distribution (how much it changes with small data perturbations) and its differential privacy.
  • Prior-Centric Privacy: The central theme is that by choosing an "informative" prior distribution that is concentrated on "smooth" regions of the model family, one can achieve differential privacy. This involves defining and analyzing conditions for "Lipschitz continuity" of the log-likelihood.
  • Posterior Sampling Mechanism: They introduce a specific algorithm (Algorithm 1: BAPS) that draws samples from the posterior. Responses to queries are then generated based on these samples, ensuring that the overall mechanism is differentially private.
  • Le Cam's Method for Lower Bounds: A novel application of Le Cam's method is used to derive lower bounds on how well an adversary can distinguish between data sets, providing a theoretical understanding of the limits of privacy.

Datasets

Not discussed in the paper. The paper focuses on theoretical frameworks and general assumptions about data distributions rather than specific datasets.

Key Insights and Ideas

  • Prior as a Privacy Lever: The most significant insight is that the prior distribution isn't just about encoding prior knowledge; it can be strategically chosen to enforce privacy guarantees. An informative prior can make the posterior less sensitive to individual data points.
  • Smoothness is Key: The paper highlights that the "smoothness" of the likelihood function (quantified by Lipschitz continuity) is crucial for achieving privacy. If the likelihood changes smoothly with parameters, small changes in data lead to small changes in the posterior.
  • Posterior Sampling as a Natural Fit: Using posterior samples for query responses is a natural extension of Bayesian inference. It allows for a flexible mechanism that can handle arbitrary queries while maintaining a fixed privacy budget.
  • Robustness and Privacy are Intertwined: The work strongly suggests that robustness and privacy are not separate goals but are deeply connected. Mechanisms that are robust to data perturbations are often also private.

Key Results

  • Differential Privacy Guarantees: The paper provides theoretical guarantees that under specific assumptions on the prior and likelihood family, the posterior distribution is differentially private.
  • Posterior Sampling Mechanism Privacy: The proposed posterior sampling mechanism (Algorithm 1) is shown to be differentially private, with privacy levels dependent on the number of samples drawn.
  • Utility-Privacy Trade-off: The authors quantify the trade-off between privacy and utility. Increasing the number of posterior samples improves utility but can decrease privacy, and vice-versa. They provide bounds on the loss of utility due to privacy.
  • Bounds on Distinguishability: They establish bounds on how easily an adversary can distinguish between two data sets based on posterior samples, showing that distinguishing requires a certain number of samples.
  • Examples: The paper demonstrates its theoretical results with examples from common distribution families like Bernoulli, Exponential, Laplace, Beta-Binomial, and Bayesian Networks, showing how the assumptions can be met.

Limitations and Future Work

  • Assumptions: The privacy guarantees rely on specific assumptions about the prior distribution and the likelihood function (e.g., Lipschitz continuity). Verifying these assumptions in practice can be challenging for complex models.
  • Constant Factors: The authors note that some constant factors in their bounds might be loose and could potentially be improved.
  • Tuning Parameters: While the number of samples offers a way to tune privacy and utility, finding the optimal balance might require further investigation.
  • Future Directions: The authors suggest exploring the trade-off between privacy and learning more deeply, improving the bounds for Assumption 2, investigating the connection between posterior sampling and the exponential mechanism, and applying these ideas to more complex models like graphical models. They also propose examining settings where adjusting both the privacy parameter ($\epsilon$) and the prior measure might be advantageous.

Summary updated

Abstract

Differential privacy formalises privacy-preserving mechanisms that provide access to a database. Can Bayesian inference be used directly to provide private access to data? The answer is yes: under certain conditions on the prior, sampling from the posterior distribution can lead to a desired level of privacy and utility. For a uniform treatment, we define differential privacy over arbitrary data set metrics, outcome spaces and distribution families. This allows us to also deal with non-i.i.d or non-tabular data sets. We then prove bounds on the sensitivity of the posterior to the data, which delivers a measure of robustness. We also show how to use posterior sampling to provide differentially private responses to queries, within a decision-theoretic framework. Finally, we provide bounds on the utility of answers to queries and on the ability of an adversary to distinguish between data sets. The latter are complemented by a novel use of Le Cam's method to obtain lower bounds on distinguishability. Our results hold for arbitrary metrics, including those for the common definition of differential privacy. For specific choices of the metric, we give a number of examples satisfying our assumptions.
BibTeX
@article{dimitrakakis2017differential,
  author = {Christos Dimitrakakis and Blaine Nelson and Zuhe Zhang and Aikaterini Mitrokotsa and Benjamin I. P. Rubinstein},
  title = {{Differential Privacy for Bayesian Inference through Posterior Sampling}},
  journal = {J. Mach. Learn. Res.},
  year = {2017},
  url = {https://openreview.net/forum?id=vohBJEu49o},
}
Figure