IncreSyn: An Efficient Framework for Incremental Differentially Private Synthetic Data

Introduction

This code base contains two construction examples of IncreSyn, constructed with two state-of-art iterative data synthesis mechanisms based on graph model, the MWEM+PGM and AIM.

For more details on Private-PGM, please visit Private-PGM.

These files also have two additional dependencies: Ektelo and autograd.

File structure

• mechanisms - Contains IncreSyn constructions which are modified for the original data synthesis of IncreSyn.
• is-mechanisms - Contains IncreSyn constructions for the incremetal data synthesis and some dataset utility tools.
• data - Contains datasets, selected cliques produced in the original data synthesis, and preferred cliques.
• src - Contains some dependencies of PGM-based mechanisms.
• UDF - Contains example UDFs.

Usage

1. Before we start, if you are only testing non-UDF parts, you could remove the pycopg2 and sqlalchemy in the requirements.txt. Moreover, when you are testing the UDF part in a Linux system like Ubuntu, you should check whether python-psycopg2 and libpq-dev are installed, or you should use apt to get them before you solve the requirements.

2. Solve the dependencies with requirements.txt. Note that we only support Python 3.

$ pip install -r requirements.txt

3. Export the src file to path. For example, in Windows, you may use:

$Env:PYTHONPATH += ";X:\IncreSyn\src"

4. Run the mechanism in mechanisms for original data synthesis, then run the corresponding mechanism under is-mechanisms.

Utility usage

• \is-mechanisms\werror.py - Calculate the workload error for a given preferred attributes file prefer.csv.

UDF usage

See README.md in /UDF.