Differential Privacy of Dirichlet Posterior Sampling

Repository to perform a private Dirichlet posterior sampling, and to reproduce the results in the following paper.

Anonymous. Differential Privacy of Dirichlet Posterior Sampling. 2022.

Run the script plots.py to generate Figure 1-4 in the paper. The code is implemented in Python 3.

The scripts were tested using Numpy 1.19.5, Scipy 1.4.1 and Matplotlib 3.3.4.

Differential Privacy

The file utils.py contains many auxiliary functions for computing the privacy guarantees stated in the paper.

• alpha2epsilon(min_alphas, gamma, Delta_2sq, Delta_inf) computes epsilon from the minimum of alpha's.
• epsilon2alpha(epsilon, gamma, Delta_2sq, Delta_inf) computes the minimum of alpha's, given a value of epsilon.
• alpha2adp(epsilon, alpha, Delta_2sq, Delta_inf) computes delta from the minimum of alpha's, given a value of epsilon.

In all of the functions above, Delta_2sq is the squared l2-sensitivity and Delta_inf is the l^{infty}-sensitivity of the sample statistics.

Private sampling

The file PrivDPS.py provides the DirichletPosteriorSampling class that tracks the privacy parameters and sample from a Dirichlet posterior. The class GaussianMechanism and LaplaceMechanism are also provided to reproduce the results in the paper.

Disclaimer

This is research code.

Licence

Apache