PyVacy: Privacy Algorithms for PyTorch

PyVacy provides custom PyTorch opimizers for conducting deep learning in a differentially private manner. Basically TensorFlow Privacy, but for PyTorch.

Getting Started

pip install pyvacy

Example Usage

import torch
from pyvacy import optim, analysis

model = torch.nn.Sequential(...)

optimizer = optim.DPSGD(
    l2_norm_clip=...,
    noise_multiplier=...,
    batch_size=...,
    lr=...,
    momentum=...,
)

epsilon = analysis.moments_accountant(
    N=...,
    batch_size=...
    noise_multiplier=...,
    epochs=...,
    delta=...,
)

for epoch in range(epochs):
    # do training as usual...

Tutorials

python tutorials/mnist.py

Training procedure achieves (3.0, 0.00001)-DP
[Epoch 1/60] [Batch 0/235] [Loss: 2.321049]
[Epoch 1/60] [Batch 10/235] [Loss: 0.952795]
[Epoch 1/60] [Batch 20/235] [Loss: 1.040896]
...