apple/pfl-research

What it solves

It provides a standardized framework for researchers to simulate private federated learning (PFL) experiments. It bridges the gap between theoretical research and actual deployments by allowing scientists to test hypotheses and run benchmarks using a unified system that supports multiple deep learning frameworks.

How it works

The framework enables the simulation of federated learning environments where models are trained across distributed data. It supports multiple levels of distributed training (processes, GPUs, and machines) and integrates privacy-preserving mechanisms, specifically local and central differential privacy. It is designed to be flexible, allowing researchers to to switch between different model types, including neural networks (via PyTorch, TensorFlow, and MLX) and Gradient Boosted Decision Trees (GBDTs).

Who it’s for

AI researchers and engineers focusing on privacy-preserving machine learning and federated learning algorithms.

Highlights

• Multi-framework support: Compatible with PyTorch, TensorFlow, and MLX.
• Privacy integration: Built-in support for local and central differential privacy.
• CONFIGURABLE simulations: Scalable simulations that can utilize multiple GPUs and machines for large experiments.
• Unified benchmarks: Provides vetted datasets and model combinations to streamline the testing of new hypotheses.