Differentially Private Bayesian Inference for Exponential Families
Authors: Garrett Bernstein, Daniel R. Sheldon
NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We design experiments to measure the calibration and utility of our method for posterior inference. We conduct experiments for the binomial model with beta prior, the multinomial model with Dirichlet prior, and the exponential model with gamma prior. The last model is unbounded and requires truncation; we set the bounds to keep the middle 95% of individuals, which is reasonable to assume known a priori for some cases, such as modeling human height. |
| Researcher Affiliation | Academia | Garrett Bernstein College of Information and Computer Sciences University of Massachusetts Amherst Amherst, MA 01002 gbernstein@cs.umass.edu Daniel Sheldon College of Information and Computer Sciences University of Massachusetts Amherst Amherst, MA 01002 sheldon@cs.umass.edu |
| Pseudocode | Yes | Algorithm 1 Gibbs Sampler, Bounded s |
| Open Source Code | No | The paper does not provide an explicit statement or link to open-source code for the methodology described. |
| Open Datasets | No | We conduct experiments for the binomial model with beta prior, the multinomial model with Dirichlet prior, and the exponential model with gamma prior. The idea is to draw iid samples (θi, xi) from the joint model p(θ)p(x | θ), and conduct posterior inference in each trial. |
| Dataset Splits | No | The paper does not provide specific training/validation/test dataset splits (e.g., percentages or sample counts). |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run experiments (e.g., GPU models, CPU types, memory). |
| Software Dependencies | No | The paper does not specify software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | We run our Gibbs sampler for 5000 iterations after 2000 burnin iterations (see supplementary material for convergence results) and We release 100 samples using the method of [12], each with ϵops = ϵ/100, such that the entire algorithm achieves ϵ-differential privacy. Private MCMC sampling [11] is a more sophisticated method to release multiple samples from a privatized posterior and could potentially make better use of the privacy budget; however, private MCMC will also necessarily be miscalibrated, and only achieves the weaker privacy guarantee of (ϵ, δ)-differential privacy for δ > 0, so would not be direct comparable to our method. OPS serves as a suitable baseline that achieves ϵ-differential privacy. We include OPS only for experiments on the binomial model, for which it requires the support of θ to be truncated to [a0, 1 a0] where a0 > 0. We set a0 = 0.1. |