FLME – Federated Learning Made Easy

FLME (Federated Learning Made Easy) is a research and development project that aims to provide a practical, scalable, and resource-aware framework for cross-device Federated Learning (FL). It is both a research platform for exploring open problems in FL (heterogeneous data, asynchronous training, scalability) and a prototype framework targeting production-level deployments at scale.

This work originates from the Bachelor Thesis of Marco Morozzi in collaboration with researchers from the University of Pisa, Massimo Torquati, and Patrizio Dazzi. You can find the thesis here: Federated Learning Made Easy (Thesis PDF)

---

✨ Motivation

Federated Learning is often centralized and synchronous: a central orchestrator coordinates clients and enforces synchronization barriers. This does not reflect the dynamics of real-world cross-device FL, where client availability is dictated by energy constraints, bandwidth, computation power, and data readiness.

FLME introduces a new paradigm called Resource Driven Federated Learning, where:

• ⚡ Clients decide when to participate (based on availability of resources such as battery, connectivity, compute, or newly collected data).
• 🔄 Training is asynchronous and fault-tolerant by design.
• 📈 The system can scale to millions of clients with low overhead.

---

🎯 Project Goals

1. An Easy Interface for FL Applications

   • Abstract away low-level implementation details.
   • Enable researchers and developers to prototype and experiment with real-world FL setups without writing orchestration logic.
   • Provide flexibility and personalization via a scripting interface (Python, Lua) to customize training strategies.

2. A High-Performance Federated Learning Engine

   • Implementation in C++ using coroutines + io_uring for highly scalable and efficient communication.
   • Fault-tolerant and scalable server backends capable of handling millions of clients.
   • Include state-of-the-art FL optimization strategies out-of-the-box.
   • Planned GPU/CUDA support for aggregation-intensive tasks.

---

🧪 Research Directions

FLME investigates fundamental challenges in Federated Learning:

• Limited scalability and flexibility of synchronous FL (stragglers, barriers, inefficiency).
• Convergence of asynchronous FL under non-IID data distributions and non-uniform client training parameters.
• Simulation and empirical validation of real-world heterogeneity.

---

🧰 Simulation Environment – flcdata/

To study FL under realistic conditions, FLME provides a simulation environment (flcdata) supporting thousands of virtual clients, where data and system heterogeneity can be controlled precisely and reproduced deterministically.

• Backend: Python
• Frontend: React
• Generates synthetic FL-oriented datasets with fine control over:
  • Label distribution skew
  • Feature distribution skew
  • Quantity skew
  • Concept drift
  • Concept shift
    (see the survey Measuring data heterogeneity in FL)

Running the Simulation Environment

cd flcdata
pip install -r requirements.txt
# then install frontend
npm install   # or yarn install / bun install
# preview mode (runs frontend + backend)
npm run runpreview

---

📂 Repository Structure

• flcdata/ → Simulation environment (frontend + backend).
• cpp/ → C++ implementation of the production-level FL framework (currently under refactoring).
• paper/ → Jupyter notebooks and scripts to reproduce results from the ongoing research paper.
• experiments/ → Exploratory implementations and replication of other FL papers.

---

🌍 Audience

• Researchers: Study asynchronous FL under reproducible heterogeneity conditions.
• Developers (future): Use a fault-tolerant and scalable framework to deploy federated applications cross-device.

---

🤝 Contributing

Contributions are welcome! 🎉
At this stage, there are no strict guidelines. If you are interested in contributing, please contact Marco Morozzi directly.

---

📖 References

• Federated Learning Made Easy (Thesis) – Marco Morozzi
• University of Pisa – Research Collaboration
• Survey on Heterogeneity in Federated Learning

---

👨‍💻 Authors & Acknowledgments

• Marco Morozzi – Developer & Researcher
• Prof. Massimo Torquati – University of Pisa
• Prof. Patrizio Dazzi – University of Pisa

---