This repository contains the code for the paper Efficient and Private Marginal Reconstruction with Local Non-Negativity, which appeared at NeurIPS 2024.
Residuals-to-Marginals (ReM) is a privacy-preserving framework for reconstructing answers to marginal queries from noisy answers to residual queries.
This repo contains the code for several marginal reconstruction algorithms based on ReM, including:
- GReM-MLE: an instantiation of ReM under Gaussian noise that uses maximum likelihood estimation to reconstruct marginals.
- GReM-LNN: an instantiation of ReM under Gaussian noise with local non-negativity constraints that uses maximum likelihood estimation to reconstruct marginals.
- EMP (Efficient Marginal Pseudoinversion): reconstructs answers to marginals from noisy answers to marginals using GReM-MLE.
This repo additionally contains two mechanisms for differentially private query answering that use the above reconstruction methods:
- ResidualPlanner: a mechanism for answering marginal queries that minimizes the noise added to the measurements.
- Scalable MWEM: a data-dependent mechanism for answering marginal queries that scales to high-dimensional datasets.
See the Demo Notebook for a demonstration of how to use these algorithms.
Outside of these reconstruction and query answering algorithms, this repo has the functionality to efficiently answer and manipulate large workloads of marginal and residual queries.
pip install git+https://github.com/bcmullins/efficient-marginal-reconstruction.git