Controlled synthetic fairness dataset for evaluating adversarial governance response (IGRA)
Description
A controlled synthetic tabular dataset (4000 instances) constructed to support fairness measurement under data-poisoning attack. Each record has two informative features drawn independently of group membership, six uninformative noise features, a binary protected-group attribute at a 65/35 split, and a binary label assigned by a median threshold on a latent score. The protected attribute is generated independently of both the predictive features and the label-generating process, so baseline group parity holds by construction. Any post-attack disparity is therefore attributable to the induced poisoning intervention and not to pre-existing imbalance in an uncontrolled source. The label split is 50/50 by construction. This file is the seed-42 instance. The generator that produces all eight seed instances used in the associated article, together with the full experimental code, is available at https://github.com/Maxwellson/igra-adversarial-ml-governance
Files
Steps to reproduce
The dataset is generated by data_generation/generate_synthetic_fairness.py in the accompanying code repository. Clone https://github.com/Maxwellson/igra-adversarial-ml-governance Install the pinned environment: pip install -r requirements.txt (NumPy 2.4.4) Run: python data_generation/generate_synthetic_fairness.py This writes data/synthetic_fairness_dataset.csv, the seed-42 instance published here. The generator is fully deterministic: it uses NumPy's default_rng(seed) and consumes random draws in a fixed order (protected group, feature_1, feature_2, the six noise features, then label noise), so the file is reproduced byte for byte on any platform with the pinned NumPy version. Other seed instances are produced by passing a different seed to make_dataset(seed). Seeds 42 through 49 are the eight instances used in the associated article. Column definitions are given in data_dictionary.csv.