Skip to content

Code to reproduce ICML 2018 paper "Differentially Private Database Release via Kernel Mean Embeddings"

License

Notifications You must be signed in to change notification settings

matejbalog/RKHS-private-database

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Differentially Private Database Release via Kernel Mean Embeddings

Matej Balog, Ilya Tolstikhin, Bernhard Schölkopf

35th International Conference on Machine Learning (ICML 2018)

[PDF] [arXiv]

This repository contains scripts to reproduce the experiments appearing in this academic paper.

Setup

Conda environment setup:

conda create -n RKHS-private-database python=3.6.3 matplotlib=2.1.0 numpy=1.13.3 pytorch=0.2.0 scikit-learn=0.19.0
source activate RKHS-private-database

Data generation

Two synthetic data files were used to generate the plots in the paper:

  • D=2: data/mixture_of_Gaussians_N100000_D2{.npz, .json}
  • D=5: data/mixture_of_Gaussians_N100000_D5{.npz, .json}

You can re-generate these files yourself by executing:

python data.py 100000 2
python data.py 100000 5

Experiments

Figure 1 ("Publishable subset" experiments)

Results of the experiments shown in Figure 1 are stored in the two files

  • D=2: results/D2_alg1_leak_M10000.json
  • D=5: results/D5_alg1_leak_M10000.json

You can re-generate these files by re-running the respective experiments as follows:

python experiments.py ../data/mixture_of_Gaussians_N100000_D2 leak --M 10000 1
python experiments.py ../data/mixture_of_Gaussians_N100000_D5 leak --M 10000 1

To then re-generate the plots shown in Figure 1, execute:

python plot.py --alg1 ../results/D2_alg1_leak_M10000.json --path_save ../figures/leaksD2
python plot.py --alg1 ../results/D5_alg1_leak_M10000.json --path_save ../figures/leaksD5
figures/leaksD2 figures/leaksD5
Figure 1 Figure 1

Figure 2 ("No publishable subset" experiments)

To re-run the experiments shown in Figure 2:

python experiments.py ../data/mixture_of_Gaussians_N100000_D2 random --M 10000 1
python experiments.py ../data/mixture_of_Gaussians_N100000_D5 random --M 10000 1
python experiments.py ../data/mixture_of_Gaussians_N100000_D2 random --M 10000 2
python experiments.py ../data/mixture_of_Gaussians_N100000_D5 random --M 10000 2

To then re-generate the plots shown in Figure 2, execute:

python plot.py --alg1 ../results/D2_alg1_random_M10000.json --alg2 ../results/D2_alg2_random_M10000.json --path_save ../figures/nodataD2
python plot.py --alg1 ../results/D5_alg1_random_M10000.json --alg2 ../results/D5_alg2_random_M10000.json --path_save ../figures/nodataD5
figures/nodataD2 figures/nodataD5
Figure 2 Figure 2

BibTeX

@inproceedings{balog2018privacy,
  author = {Balog, Matej and Tolstikhin, Ilya and Sch\"olkopf, Bernhard},
  title = {Differentially {Private} {Database} {Release} via {Kernel} {Mean} {Embeddings}},
  booktitle = {35th International Conference on Machine Learning (ICML)},
  year = {2018},
  month = {July}
}

About

Code to reproduce ICML 2018 paper "Differentially Private Database Release via Kernel Mean Embeddings"

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published