Differentially private synthetic tabular data with Private Evolution (PE). Each script downloads a public dataset (train/test split + metadata), runs PE, and writes results to results/tabular/<experiment>/:
synthetic_tab/— generated synthetic CSVscheckpoint/— per-iteration checkpoints (runs resume from here)- classifier accuracy (
TabClassifier) and Wasserstein-style marginal distance (ComputeWSD) logged tolog.txtand CSV
Our code automatically downloads the datasets, which are available at https://github.com/toan-vt/cloud-data-store/tree/main/tabular.
pip install private-evolution[tabular]Or, for an editable install from the repo root:
pip install -e ".[tabular]"XOR datasets with a configurable number of features (from 1 to 7):
python xor_stress_test.py --num-features 1
python xor_stress_test.py --num-features 2Structural causal model (SCM) data with a selectable prior function:
python scm.py --prior-function rff # choices: tree, nn, rffpython artificial_characters.py
python person_activity.py
python adult.py
python breast_cancer.py