aia-fairness 0.0.2

Various attributes inferences attacks tested against fairness enforcing mechanisms

AIA_fairness

Code repository of the article. The repositries segments the pipeline as to isolate key componants of the experimental analysis.

Installation and dependencies

(TODO) By using pypi : pip install aia_fairness

Or clone the repository then run pip install --editable . in the directory containing pyproject.toml

All dependencies are specified in pyproject.toml and installed automatically by pip.

How to use

Global configuration parameters can be set by editing config.py

Dataset automatic download and processing

Part of the dataset uses the kaggle API to download the data. Hence you need to include your API key in ~/.kaggle/kaggle.json.

aia_fairness provides automatic download, formatic and saving of the dataset used in the article. To use this feature and thus saving data in the data_format directory simply run once

import aia_fairness.dataset_processing as dp
dp.fetch()

Then you can load any dataset easily form anywhere in your code with

import aia_fairness.dataset_processing as dp
data = dp.load_format(dset, attrib)

You can refer to the full implementation in test/target_training.py.

Running all the expriments form the paper

** Heavy computing power required **

One you have downloaded all the data you can run all the expirement of the article by running

python -m aia_fairness.experimental_stack

or import aia_fairness.experimental_stack in the python interpreter.

Training a target model

aia_fairness.models.target contains various target model type. The target models available are :

• RandomForest
• RandomForest_EGD fairlearn is used to impose EGD
• NeuralNetwork
• NeuralNetwork_Fairgrad Original implementation of the fairgrad paper
• NeuralNetwork_AdversarialDebiasing Uses the fairlearn implementation of Adversarial Debisaing

For instance to train a random forest (based on sklearn) you can

import aia_fairness.models.target as targets
T = dp.split(data,0)
target.fit(T["train"]["x"], T["train"]["y"])
yhat = target.predict(T["test"]["x"])

Evaluation of a target model

aia_fairness.evaluation provides the metrics used in the article.

import aia_fairness.evaluation as evaluations
utility = evaluations.utility()
fairness = evaluations.fairness()
utility.add_result(T["test"]["y"], yhat, T["test"]["z"])
fairness.add_result(T["test"]["y"], yhat, T["test"]["z"])
utility.save(type(target).__name__,dset,attrib)
fairness.save(type(target).__name__,dset,attrib)

The save method, as called in the example, creates a directory structure of the form :

#|result/target/<target type>
#|  |<name of the metric class>
#|  |   |<dset>
#|  |   |   |<attrib>
#|  |   |   |   <Name of metric 1>.pickle
#|  |   |   |   <Name of metric 2>.pickle
#|  |   |   |   <Name of metric ..>.pickle
#|  |<name of another metric class>
#|  |   |<dset>
#|  |   |   |<attrib>
#|  |   |   |   <Name of another metric 1>.pickle
#|  |   |   |   <Name of another metric 2>.pickle
#|  |   |   |   <Name of another metric ..>.pickle

Training an attack

aia_fairness.models provides the two types of AIA attack described in the paper:

• Classification for hard labels
• Regression for soft labels

from aia_fairness.models import attack as attacks
aux = {"y":T["test"]["y"],
       "z":T["test"]["z"],
       "yhat":yhat}
aux_split = dp.split(aux,0)
classif = attacks.Classification()
classif.fit(aux_split["train"]["yhat"], aux_split["train"]["z"])

Evaluation of an attack

Similarly with evaluating the target model aia_fairness.evaluation.attack can be used to save the accuracy and the balanced accuracy of the attack.

Plots and graphs

TODO

Tests scripts

Various tests are provided in the test directory.

• download_data.py Fetches all the data form the diferent sources (don't forget to set your kaggle API key)
• target_training.py Loads a dataset, split it with 5-folding (cross validation), train a target model on the data, and computes the metrics for this target model

Directories structure

• data_processing contains code that downloads, preprocess and saves the dataset in a uniformed pickle format exploitable by the rest of the pipeline using load_format(dataset,attribute) function of the utils.py file.