Rao Differential Privacy

Carlos Soto

Abstract

Rao Differential Privacy (Rao DP) is a novel privacy framework grounded in information geometry. By utilizing the Rao distance—a proper metric on the manifold of probability distributions—instead of traditional statistical divergences, this framework provides a geometric interpretation of privacy. A key advantage of Rao DP is its "tighter" composition rule, where privacy loss across multiple analyses combines via the $\ell_2$-norm (Pythagorean sum) rather than the standard $\ell_1$-norm (linear sum), significantly preserving data utility without compromising privacy guarantees.

Introduction

Differential Privacy (DP) has become the gold standard for statistical data release with individual identity protection. However, existing variants—such as pure DP, Rényi DP, or Zero-concentrated DP—rely on statistical divergences (e.g., Kullback-Leibler or Rényi divergence). These divergences are not true metrics and often fail the triangle inequality. This mathematical property leads to "budget inflation" during composition, where the privacy loss of multiple queries typically sums linearly (the $\ell_1$ sum). Such rapid budget depletion can render datasets useless for complex analysis. This paper proposes that this inflation is an artifact of using divergences and introduces Rao DP as a distance-based alternative.

Geometric Foundations of Rao DP

The proposed framework shifts the perspective from abstract functions to information geometry, treating probability distributions as points on a curved Riemannian manifold.

  1. Fisher Information Metric: The Fisher Information Matrix serves as a local metric to quantify how a distribution changes relative to parameter shifts.
  2. Rao Distance: Privacy loss is defined as the geodesic distance—the shortest path on the curved manifold—between the output distributions of a mechanism on adjacent datasets $D$ and $D'$.
  3. Definition: A mechanism satisfies $\theta$-Rao DP if the Rao distance between these distributions is bounded by $\theta$.

This approach ensures immunity to post-processing and facilitates a more natural derivation of composition rules through geometric properties.

Key Insights

A primary contribution is the Pythagorean Composition. On a product manifold, distances combine according to the Pythagorean theorem. Consequently, for two analyses with budgets $\theta_1$ and $\theta_2$, the total privacy cost is $\sqrt{\theta_1^2 + \theta_2^2}$. This $\ell_2$ composition is significantly tighter than the $\ell_1$ sum ($\theta_1 + \theta_2$) used in standard DP variants.

Furthermore, Rao DP provides a unified view of Gaussian Differential Privacy (GDP). While GDP achieves similar composition behavior through complex hypothesis testing trade-off functions, Rao DP arrives at the same result through pure geometry. It also eliminates the need for the $\delta$ parameter found in "Approximate DP," which represents the probability of total privacy failure, by handling noise naturally within the single $\theta$ parameter.

Results and Comparison

The paper formally proves the $\ell_2$-norm composition rule: $\text{Total Budget} = \sqrt{\sum \theta_i^2}$. It also establishes compatibility with standard noise-adding mechanisms, including the Laplace, Gaussian, and Generalized Gaussian mechanisms. For the Laplace mechanism, $\theta$-Rao DP is equivalent to $\epsilon$-pure DP when $\theta = \epsilon$, yet it benefits from the superior $\ell_2$ composition properties.

Comparison of privacy frameworks, their budget parameters, and composition rules.

Limitations and Future Work

While Rao DP offers a robust theoretical framework, several challenges remain. The current formulation focuses on continuous probability densities; formal results for discrete distributions are yet to be established. Additionally, computing Rao distances for complex, multi-parameter distributions (e.g., multivariate Gaussians with varying covariance) often lacks closed-form solutions, presenting a computational bottleneck. Future research may extend this framework to non-Euclidean data spaces like graphs or more complex manifolds.

Summary updated

Abstract

Differential privacy (DP) has recently emerged as a definition of privacy to release private estimates. DP calibrates noise to be on the order of an individual's contribution. Due to this calibration, a private estimate obscures any individual while preserving the utility of the estimate. Since the original definition, many alternate definitions have been proposed. These alternates have been proposed for various reasons including improvements on composition results, relaxations, and formalizations. Nevertheless, thus far nearly all definitions of privacy have used a divergence of densities as the basis of the definition. In this paper we take an information geometry perspective towards differential privacy. Specifically, rather than define privacy via a divergence, we define privacy via the Rao distance. We show that our proposed definition of privacy shares the interpretation of previous definitions of privacy while improving on sequential composition.
BibTeX
@misc{soto2025rao,
  author = {Carlos Soto},
  title = {{Rao Differential Privacy}},
  year = {2025},
  eprint = {2508.17135},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.17135},
}
Figure