Rényi Pufferfish Privacy Based on Gaussian Priors: From a Single Gaussian to Mixture Models
Rényi Pufferfish Privacy with Gaussian-based Priors: From Single Gaussian to Mixture Model
The researchers propose noise mechanisms based on Gaussian priors for the relevant data privacy framework Rényi Pufferfish Privacy (RPP), reducing the average required noise by 48.9% compared with a recent RPP additive-noise baseline on three UCI datasets. For a single Gaussian prior, the paper derives the exact Rényi divergence after Gaussian perturbation and gives a relaxed closed-form sufficient condition for (α,ε)-RPP; for non-Gaussian, multimodal priors, it approximates them with Gaussian mixture models and gives a sufficient condition via optimal transport. These mechanisms require less noise under both statistical (RAW, MEAN) and model-output (BNN, GP) queries, improving the privacy-utility trade-off of RPP.