Impact of Multi-dimensional Local Differential Privacy on Fairness

Repository for the paper: On the Impact of Multi-dimensional Local Differential Privacy on Fairness.

Codes & Datasets

All the experiments are implemented in Python 3. We use Random Forest model for classification with its default hyper-parameters and we use the ten-fold cross-validation technique. For k-RR mechanism, we use the implementation in Multi-Freq-LDPy Python library. Since LDP protocols and ML algorithms are randomized, we report average results over 20 runs.

• The datasets folder includes all the used and generated datasets
• The Results folder contains all the results (as csv files) of fairness metrics before and after applying the KRR mechanism settings for all the datasets. The settings applied in this study are:
  • sLDP: only the protected attribute is obfuscated.
  • allsLDP: a list of sensitive attributes (including the protected attribute) is obfuscated using KRR independently by applying the k-based privacy splitting solution.
  • combLDP: a list of sensitive attributes (including the protected attribute) is obfuscated using the combined setting of KRR.
• For each dataset, three Jupyter notebooks are available:
  • 1_Generated_data.ipynb: Jupyter notebook for generating (for the synthetic dataset) or preprocessing (for the Adult and Compas datasets) the data.
  • 2_Experiments.ipynb: Jupyter notebook for computing fairness metrics and accuracy before and after applying KRR mechanism settings.
  • 3_Generating_Plots.ipynb: Jupyter notebook for generating all the results of the paper.

Environment

Our codes were developed using Python 3 with numpy, and pandas libraries. The versions are listed below:

• Python 3.7.3
• Numpy 1.21.2
• Pandas 1.3.3
• Multi-freq-ldpy 0.2.4

On-going

The repository is still under cleaning/generalization of the codes and the documentation.

Merits

• We use the randomized mechanism KRR implemented in the multi-freq-ldpy library.
• We use the Adult dataset from the folktables library.
• We use the Compas dataset from the Kaggle repository.

Contact

For any questions, please contact:

Karima Makhouf: karima.makhlouf [at]inria.fr

Héber H. Arcolezi: heber.hwang-arcolezi [at] inria.fr

Sami Zhioua: zhioua[at]lix.polytechnique.fr