Geometric Renyi Differential Privacy: Ricci Curvature Characterized by Heat Diffusion Mechanisms

A team of researchers has introduced a groundbreaking new framework for differential privacy (DP) called 'Geometric Rényi Differential Privacy.' The work, led by Xiaotian Chang, Yangdi Jiang, and Cyrus Mostajeran, fundamentally rethinks how to add privacy-preserving noise to data that isn't flat but exists on curved, high-dimensional spaces known as Riemannian manifolds. Examples of such data include medical brain scans, climate sensor networks, or complex financial models. The core innovation is establishing a direct mathematical link between the strength of a privacy guarantee (Rényi DP) and the geometric property of the data's underlying space, specifically its Ricci curvature.

For data on manifolds with nonnegative Ricci curvature, the team proposes a highly efficient mechanism based on the mathematics of heat diffusion—imagine privacy noise spreading like warmth across a curved surface. For more general, arbitrarily curved data spaces, they developed an alternative approach using Langevin stochastic processes. Both methods offer a continuous trade-off between privacy and data utility and crucially avoid the computationally expensive step of 'normalization' required by many existing DP techniques. The researchers demonstrated the framework's power by applying it to privately estimate a 'generalized Fréchet mean'—a complex, nonlinear average on a curved space—and showed superior performance over prior methods in numerical experiments.