Federated learning (FL) is an emerging machine learning training paradigm in which clients own their data and collaborate to train a global model without revealing any data to the server or other participants.
Researchers often conduct experiments in a simulation environment to quickly iterate ideas. However, existing open source tools do not offer the efficiency needed to simulate FL on larger, more realistic FL data sets. Introducing pfl-research, a fast, modular, and easy-to-use Python framework for simulating FL. It supports TensorFlow, PyTorch, and non-neural network models, and is tightly integrated with state-of-the-art privacy algorithms.
We study the speed of open source FL frameworks and show that pfl-research is 7 to 72 times faster than alternative open source frameworks in common cross-device configurations. This acceleration will significantly increase the productivity of the FL research community and allow hypotheses to be tested on realistic FL data sets that were previously too resource-intensive. We released a set of benchmarks that evaluate the overall performance of an algorithm in a diverse set of realistic scenarios.
