ldp-audit

A tool for auditing Locally Differentially Private (LDP) protocols. LDP-Audit is a work in progress, and we expect to release new versions frequently. The tool and framework accompanies our PETS 2024 paper "Revealing the True Cost of Locally Differentially Private Protocols: An Auditing Perspective ":

@article{arcolezi2024revealing,
  title={Revealing the True Cost of Locally Differentially Private Protocols: An Auditing Perspective},
  author={Arcolezi, H{\'e}ber H and Gambs, S{\'e}bastien},
  journal={Proceedings on Privacy Enhancing Technologies},
  volume={2024},
  number={4},
  pages={123--141},
  year={2024},
  doi = {10.56553/popets-2024-0110},
}

Installation

To install the dependencies, you can use the following commands:

git clone https://github.com/hharcolezi/ldp-audit.git
cd ldp-audit
pip install -r requirements.txt

Usage

To reproduce the experiments from the paper, run:

python experiment_1.py
python experiment_2.py (depend on files generated by experiment_1.py)
python experiment_3.py
python experiment_4.py
python experiment_5.py
python experiment_6.py

Example:

# Import necessary modules from ldp-audit package
from ldp_audit.base_auditor import LDPAuditor

# Define audit parameters
seed = 42 # for reproducibility
nb_trials = int(1e6) # trial count should be high, e.g., >=10^6
alpha = 1e-2 # confidence level should be small, e.g., 0.05, 0.01
epsilon = 1 # the theoretical privacy parameter
k = 2 # the domain size

print('=====Auditing pure LDP protocols=====')
delta = 0.0
pure_ldp_protocols = ['GRR', 'SS', 'SUE', 'OUE', 'THE', 'SHE', 'BLH', 'OLH']
auditor_pure_ldp = LDP_Auditor(nb_trials=nb_trials, alpha=alpha, epsilon=epsilon, delta=delta, k=k, random_state=seed, n_jobs=-1)

for protocol in pure_ldp_protocols:
    eps_emp = auditor_pure_ldp.run_audit(protocol)
    print("{} eps_emp:".format(protocol), eps_emp)

print('\n=====Auditing approximate LDP protocols=====')
delta = 1e-5
approx_ldp_protocols = ['AGRR', 'ASUE', 'AGM', 'GM', 'ABLH', 'AOLH']
auditor_approx_ldp = LDP_Auditor(nb_trials=nb_trials, alpha=alpha, epsilon=epsilon, delta=delta, k=k, random_state=seed, n_jobs=-1)

for protocol in approx_ldp_protocols:
    eps_emp = auditor_approx_ldp.run_audit(protocol)
    print("{} eps_emp:".format(protocol), eps_emp)

To Do

• Deploy the tool as a Python package to PyPI for easier installation and integration.
• Add more detailed documentation and examples.
• Expand the testing suite for better coverage.
• Optimize performance for large-scale data.

Contributions

We welcome contributions! Please fork the repository and submit pull requests.

Contact

For any question, please contact Héber H. Arcolezi: heber.hwang-arcolezi [at] inria.fr

License

This project is licensed under the MIT License. See the LICENSE file for details.