deepee

deepee is a library for differentially private deep learning in PyTorch. More precisely, deepee implements the Differentially Private Stochastic Gradient Descent (DP-SGD) algorithm originally described by Abadi et al.. Despite the name, deepee works with any (first order) optimizer, including Adam, AdaGrad, etc.

It wraps a regular PyTorch model and takes care of calculating per-sample gradients, clipping, noising and accumulating gradients with an API which closely mimics the PyTorch API of the original model.

Check out the documentation here

For paper readers

If you would like to reproduce the results from our paper, please go here