Individual privacy accounting for DP-SGD

This repo implements the algorithm in Individual Privacy Accounting for Differentially Private Stochastic Gradient Descent for training ResNet-20 on CIFAR-10.

Build the environment

pip install -r requirements.txt

Run DP-SGD with individual privacy accounting

This command trains a resnet20 model on CIFAR-10. The training takes ~1.5 hours with a single A100 GPU.

python main.py --private --sess example_exp --sigma 2.2 --n_epoch 200 --clip 15

Visualize individual privacy parameters

After training, you can visualize the histogram of individual privacy and estimation errors. The figures are saved in the figs folder.

python visualization.py --sess example_exp

Figures