Improved DP-SGD for differentially private learning. Codes for the experiment part of our paper "Improving the Gradient Perturbation Approach for Differentially Private Optimization" at NeurIPS 2018 ppml workshop https://ppml-workshop.github.io/ppml/.
Step1:
Install python3
Install pytorch at https://pytorch.org/
Step2:
Navigate to this repository's directory and install required packages.
pip install -r requirements.txt
Step3:
Run the code, for example:
python main.py --dataset mnist --epsilon=2.0 --delta=1e-5 --momentum=0.6 --SGN=2
runs our mnist experiment on convex objective using momentum and varying clip bound.