/dp-finder

Differential Privacy Testing System

Primary LanguagePythonOtherNOASSERTION

DP-Finder

DP-Finder is a system that automatically derives lower bounds on the differential privacy enforced by algorithms.

This repository contains the code used for the experiments in the paper https://www.sri.inf.ethz.ch/publications/bichsel18dpfinder.

Setup

To install DP-Finder, you can use docker or install it locally. Before starting, clone this repository and navigate to the directory of this README file.

Docker

For a simple (but less efficient) setup using docker, run

$ cd dp-finder/docker
$ make launch # build and run the docker image
[sudo] password for user: *************

Locally

Alternatively, you can set up DP-Finder on your local machine. The main requirements are

$ sudo apt-get install python3 python3-pip python3-tk libboost-all-dev
[...]
$ pip3 install virtualenv
[...]
$ sudo apt-get install texlive-full
[...]

In addition, the DP-Finder requires PSI (to confirm found violations exactly). Make sure that PSI can be accessed by running psi, by adding it to /usr/local/bin. For example, assuming that psi was installed to /opt/psi/psi, run:

ln -s -f "/opt/psi/psi" "/usr/local/bin"

See the Dockerfile for all relevant packages on for how to install PSI. In case of issues with the setup, also see the preparation script (which is automatically run when you run ./test.sh or ./run.sh), which contains some optional commands that may fix your errors.

Getting Started

Before running any commands, prepare the environment (sets up the PYTHONPATH, compiles dependencies, etc) by running

root@febda6ac8a18:/implementation# source ./prepare.sh

Testing the build

To run the unittests of DP-Finder, run

root@febda6ac8a18:/implementation# ./test.sh

This should take around 10 minutes.

Finding lower bounds

To find lower bounds for the encoded algorithms, run

root@febda6ac8a18:/implementation# ./run.sh

The runner generates plots, which are saved in ./dpfinder/log_parser/figures. The runtime of this script depends on your machine (expect <1day).

Running DP-Finder on a single algorithm

root@febda6ac8a18:/implementation# python3 dpfinder/searcher/search.py --alg aboveThreshold

To get a description of all parameters to dp-finder, run

root@febda6ac8a18:/implementation# python3 dpfinder/searcher/search.py --help

Adding more algorithms

To add a new algorithm newAlg for testing:

Citing This Framework

@inproceedings{Bichsel:2018:DFD:3243734.3243863,
 author = {Bichsel, Benjamin and Gehr, Timon and Drachsler-Cohen, Dana and Tsankov, Petar and Vechev, Martin},
 title = {DP-Finder: Finding Differential Privacy Violations by Sampling and Optimization},
 booktitle = {Proceedings of the 2018 ACM SIGSAC Conference on Computer and Communications Security},
 series = {CCS '18},
 year = {2018},
 isbn = {978-1-4503-5693-0},
 location = {Toronto, Canada},
 pages = {508--524},
 numpages = {17},
 url = {http://doi.acm.org/10.1145/3243734.3243863},
 doi = {10.1145/3243734.3243863},
 acmid = {3243863},
 publisher = {ACM},
 address = {New York, NY, USA},
 keywords = {differential privacy, lower bounds, optimization, sampling},
}

Contributors

License and Copyright