domias 0.0.1

DOMIAS, a density-based MIA model that aims to infer membership by targeting local overfitting of the generative model.

DOMIAS: Membership Inference Attacks against Synthetic Data through Overfitting Detection

:rocket: Installation

The library can be installed from PyPI using

$ pip install domias

or from source, using

$ pip install .

Experiments

1. Experiments main paper

To reproduce results for DOMIAS, baselines, and ablated models, run

cd experiments
python3 domias_main.py --seed 0 --gan_method TVAE --dataset housing --training_size_list 30 50 100 300 500 1000 --held_out_size_list 10000 --gen_size_list 10000 --training_epoch_list 2000

changing arguments training_size_list, held_out_size_list, gen_size_list, and training_epoch_list for specific experiments over ranges (Experiments 5.1 and 5.2, see Appendix A for details) and gan_method for generative model of interest.

or equivalently, run

cd experiments && bash run_tabular.sh

2. Experiments no prior knowledge (Appendix D)

If using prior knowledge (i.e., no reference dataset setting), add

--density_estimator prior

3. Experiment images (Appendix B.3)

Note: The CelebA dataset must be available in the experiments/data folder.

To run experiment with the CelebA dataset, first run

cd experiments && python3 celeba_gen.py --seed 0 --training_size 4000

and then

cd experiments && python3 celeba_eval.py --seed 0 --training_size 4000

:hammer: Tests

Install the testing dependencies using

pip install .[testing]

The tests can be executed using

pytest -vsx

Citing

TODO