Impact of using synthetic data on MIA and AIA

API documentation

The documentation is generated using sphinx. To view it, clone the repo and use your web browser to open docs/build/html/index.html. To regenerate the API documentation:

sphinx-apidoc -f -F -o docs src;
sphinx-build -M html docs docs/build/

Tests

Unit tests are setup in the test directory. They use the python unittest packages. To run all tests:

python -m unittest discover -v -p "*.py" -s test

Installation

Instal with pip in a venv. In the directory containing pyproject.toml :

 pip install --editable .

Usage

Running experiments through the command line interface

python -m synthetic_aia_mia.experiments.adult <generator> <k>;

• Generator is either identity or ctgan.
• k folding step in 0,1,2,3,4.

Key features

#Load data
from synthetic_aia_mia.fetch_data import adult
from synthetic_aia_mia.fetch_data import utk
data_adult = adult.load()
data_utk = utk.load()

#Train a neural network an adult 
form synthetic.predictor.adult import AdultNN
adultNN = AdultNN()
adultNN.fit(data_adult["train"])

#Evalute trained neural network
adultNN.predict(data_adult["test"])

#Generate synthetic data
from synthetic_aia_mia.generator.adult import ctgan 
from synthetic_aia_mia.generator import identity 
synthetic_identity = identity(data_adult(["train"]))
synthetic_ctgan = ctgan(data_adult(["train"]))

Datasets

Adult

We are using folktables adult.

UTKFaces

From Kaggle: jangedoo/utkfaces-new. For loading and parsing files we use aia_fairness.dataset_processing