nfl

This project uses NIllion to perform a federated learning training procedure for neural networks. It provides a framework for arbitrary networks once Nillion scales to more computations.

It provides a script to autogenerate a program that can perform federated learning updates for a model as well as functionality to store local weights as secrets, share computation permissions for these secrets with the global party, and then update the local weights at every training epoch.

This project is entirely using Nillion and Python. Pytorch was the deep learning framework used, as well as a few other utility libraries in Python.

Notable things include autogenerated Nillion program for arbitrary neural networks. This allows the approach to be scalable to many more complex architectures. It also includes the entire end to end training loop in a series of scripts.