We present here a very simple example of combining Federated Learning (FL) with Differential Privacy (DP), which can be an interesting baseline to experiment with these great technologies. More specifically, we show how the DP Opacus library released by PyTorch can be used in PySyft FL workflows with very little overhead.

We label encrypted images with an encrypted ResNet-18 using PySyft and Function Secret Sharing.

We use the PySyft library to encrypt a neural network and privately classify MNIST images using Secure Multi-Party Computation (SMPC). We achieve classification in <33ms with>98% accuracy over local (virtualized) computation.

We use the PySyft library to train a PyTorch neural network on MNIST using Secure Multi-Party Computation (SMPC). We combine PyTorch nets, SMPC & Autograd in a single demo.

We show how to do Federated Learning with PySyft in a very easy way by training a Convolutional Neural Network over a distributed version of the MNIST Dataset.

Federated Learning

Federated Learning is a very exciting and upsurging Machine Learning technique for learning on decentralized data. The core idea is that a training dataset can remain in the hands of its producers (also known as _workers_) which helps improve privacy and ownership, while the model is shared between workers. One popular application of Federated Learning is for learning the "next word prediction" model on your mobile phone when you write SMS messages: you don't want the data used for training that predictor — i.e. your text messages — to be sent to a central server. (Read more.)