Private Rate-Constrained Optimization

dpraco is a research codebase for training classifiers with differential privacy and fairness constraints. The examples in this repository focus on running DpRaco with tabular datasets and inspecting the privacy/fairness trade-offs.

What is in this repository

• src/dpraco/: core implementation (data loaders, models, algorithms, training loop, accounting).
• examples/: notebook demos that run end-to-end experiments.
• examples/conf/: composable OmegaConf/Hydra-style configuration files used by the demos.
• data/: local dataset files used by the notebooks.

Setup

This project uses uv.

uv sync --dev

Optional shell activation:

source .venv/bin/activate

You can also run commands without activating the environment by prefixing them with uv run.

Examples

The examples/ folder contains three notebooks:

• examples/dpraco_adult_demographic_parity_demo.ipynb
  Runs DPRACO on Adult with a Demographic Parity constraint.
• examples/dpraco_folktables_demographic_parity_demo.ipynb
  Runs DPRACO on Folktables with a Demographic Parity constraint.
• examples/dpraco_heart_false_negative_rate_demo.ipynb
  Runs DPRACO on Heart with a False Negative Rate constraint.

Important: launch notebooks from examples/

The notebooks resolve config paths relative to the current working directory (they expect conf/ to be in that directory). Start Jupyter from examples/:

cd examples
uv run jupyter lab

Then open one of the dpraco_*_demo.ipynb notebooks.

How each demo is configured

Each notebook builds a runtime config by merging:

1. base config: examples/conf/dpraco.yaml
2. dataset config: examples/conf/dataset/*.yaml
3. algorithm config: examples/conf/algorithm/*.yaml
4. model config: examples/conf/model/logistic_regression.yaml
5. demo overrides: examples/conf/demo/*.yaml

The demo override files used by the notebooks are:

• examples/conf/demo/adult_demographic_parity_demo.yaml
• examples/conf/demo/folktables_demographic_parity_demo.yaml
• examples/conf/demo/heart_false_negative_rate_demo.yaml

In the notebooks, the most common parameters to tune are:

• cfg.training_params.target_epsilon
• cfg.algorithm.gamma
• cfg.training_params.number_of_steps

What the notebooks do

All three demos follow the same flow:

1. load and merge config files with OmegaConf
2. build data streams via dpraco.data.get_data_stream
3. initialize a model via dpraco.models.get_model
4. run training/evaluation with dpraco.training.training_routine.train_and_evaluate
5. inspect and plot accuracy and fairness metrics over training steps

Generated metrics include columns such as:

• train_accuracy, val_accuracy, test_accuracy
• train_soft_constraint, train_hard_constraint (and corresponding val/test columns)
• per-classification-report values (for example recall, precision, f1-score)

Notes

• The notebooks currently set JAX_CPU_ONLY = True, so they default to CPU execution unless you modify that setup cell.
• Data paths in configs point to ../data/... relative to examples/, so local datasets should be present under the repository data/ directory.

Cite us

If you use our library or build on top of our work, please consider citing us:

@inproceedings{
  yaghini2026private,
  title={Private Rate-Constrained Optimization with Applications to Fair Learning},
  author={Mohammad Yaghini and Tudor Cebere and Michael Menart and Aur{\'e}lien Bellet and Nicolas Papernot},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=mex3rvs2KX}
}