Differentially Private Covariance Revisited

This repo contains the official code for the implementations of the algorithms and experiments in the paper "Differentially Private Covariance Revisited".

Folder	Description
adaptive	implementations for our algorithms and the Gaussian mechanism
exponential	implementation for Exponential mechanism for covariance estimation
coinpress	a copy of the code from https://github.com/twistedcubic/coin-press, containing implementation of coinpress

Requirements

The algorithms and experiments were implemented in Python (v3.9).

The following packages are required and can be installed via conda or pip:
pytorch v1.9.0
numpy v1.22.3
scipy v1.8.0
scikit-learn v1.0.1

conda install numpy==1.22.3
conda install pytorch==1.9.0 torchvision==0.10.0 torchaudio==0.9.0 cpuonly -c pytorch
conda install scipy==1.8.0
conda install scikit-learn==1.0.1

Evaluation

To reproduce the results in a specific figure, run its corresponding script listed below.

File name	Corresponding figure	Description
test_d_N1.py	Figure 1.	Synthetic dataset test, fixing tr=1
test_Ns.py	Figure 2(a).	Synthetic dataset test, various tr
test_d.py	Figure 2(b).	Synthetic dataset test, various d
test_n.py	Figure 2(c).	Synthetic dataset test, various n
test_rho.py	Figure 2(d).	Synthetic dataset test, various rho
test_mnist_fix_n.py	Figure 3.	MNIST dataset test
test_news_rho_l2.py	Figure 4(a)-(c).	News dataset test, unit l2 norm
test_news_rho_maxl2.py	Figure 4(d)-(e).	News dataset test, normalized by max l2 norm
test_mnist_digit.py	Figure 5.	MNIST dataset individual digits test
test_d_N1_pure.py	Figure 6.	Synthetic dataset test, fixing tr=1 (pure DP)
test_Ns_pure.py	Figure 7(a), 7(d)	Synthetic dataset test, various tr (pure DP)
test_d_pure.py	Figure 7(b), 7(e)	Synthetic dataset test, various d (pure DP)
test_eps_pure.py	Figure 7(c), 7(f)	Synthetic dataset test, various epsilon (pure DP)

For example, to run the test in Figure 1:

python test_d_N1.py

Note: when running the corresponding scripts for the tests on real-world datasets, the code will check whether the datasets are in the directory "./data/".
If not, it will download them using the links embedded in the code to that directory.

Compute Resources

All experiments reported in the paper were run on a linux machine with 48 CPUs and 256GB RAM.