This repository contains the full implementation of the Relaxed Adaptive Projection (RAP) mechanism used in the experiments of the Pushing the Boundaries of Private, Large-Scale Query Answering paper by Brendan Avent and Aleksandra Korolova.
This project was developed to primarily to run on systems with Python 3.6+ that have an Nvidia GPU. It is assumed that the relevant CUDA and CuDNN libraries are installed and compatible with both the GPU and the (to-be-installed) JAX version.
Installing the Python packages required by this implementation is done by executing the pip command:
pip install -r requirements.txt
To answer very large sets of queries (> 2*10^9), JAX must be set to 64-bit mode on the underlying system as described here. On Linux from a BASH shell, a simple way to set this permanently is to execute the following command:
echo 'export JAX_ENABLE_X64=True' >> ~/.bashrc
The datasets evaluated in the paper are the ADULT and LOANS datasets.
These datasets can be automatically downloaded by running the downloader.py file in the data directory. Alternatively, the CSVs can be manually downloaded and placed in their corresponding subdirectories within the data directory.
Finally, ensure that the root of this project is in the system's PYTHONPATH.
This implementation is split between two primary directories at the project's root: data and rap.
The data directory contains files related to downloading, importing, and managing, and onehot encoding/decoding the relevant datasets.
The rap directory contains files for managing individual r-of-k thresholds, managing workloads of r-of-k thresholds, executing the RAP mechanism, and evaluating/storing the results of the RAP mechanism and the other simple baseline mechanisms.
example.py is a simple example script that illustrates how to programmatically set up and run a single simulation of the RAP mechanism.