SensitiveAttributesReconstructionCorrector

This repository contains the implementation of the Reconstruction Corrector component introduced in our article "Exploiting Fairness to Enhance Sensitive Attributes Reconstruction"

File reconstruction_tools_fairlearncomputation.py contains the implementation of our contributions.

File example_reconstruction_correction.py provides an example use of our Python modules. More precisely, it performs the following tasks:

• It loads the UCI Adult Income dataset
• It trains a fair model using this dataset, and the ExponentiatedGradient fair learning technique from the FairLearn library.
• It trains a baseline adversary (corresponding to A' in our paper) using our MachineLearningAttacker object.
• It performs the adversary's confidence scores exponentiation process using our scale_probas_manual function.
• It performs the reconstruction correction using our ReconstructionCorrector object.
• It evaluates its quality (and the improvement brought by the correction step) using our evaluate_reconstruction function.

Requirements

• Our ReconstructionCorrector object implements the reconstruction correction method described in our article. It builds and solves a Constraint Programming (CP) model, using the Docplex Python modelling tool and the IBM CP Optimizer solver. Both are required for the ReconstructionCorrector to be able to run.
• Our example script uses the ExponentiatedGradient fair learning technique from the FairLearn library. To run it, FairLearn must be installed.
• Our example script uses several tools from the scikit-learn library. To run it, scikit-learn must be installed.
• Other popular required libraries: numpy, pandas

Example output

Hereafter is an example output produced by our example script for the statistical parity metric with unfairness tolerance epsilon = 0.0 (random_state_value=42):

Data is split among: 
Training set: 15375 examples
Attack set: 14923 examples
Test set: 14924 examples
Number of features for inference:  13 

Training target fair model...
Trained!
Train unfairness (statistical_parity) =  1.333702367489753e-06
Train accuracy:  0.8394146341463414
Test accuracy:  0.8308094344679711 

Training baseline adversary...
Machine learning attacker ready. Accuracy on attack set is  0.7639884741673926
Trained!
Computing the probabilities exponentiation factor value...
Exponentiating all probabilities to factor  6 

Performing the reconstruction correction...
Done! 

Corrected reconstruction:  0.8300487804878048  (was  0.7654634146341464  before)
Absolute improvement =  0.06458536585365848
Relative improvement =  0.08437420341575318