Federated Learning Platform

Production-grade federated learning with FedAvg, FedProx, differential privacy, and secure aggregation.

---

Quick Start (One Command)

./launch-platform.sh

Then visit:

• Dashboard: http://localhost:8050
• MLflow: http://localhost:5000

That's it. The platform will:

1. Start FL server and clients
2. Train across multiple clients
3. Show real-time progress
4. Track experiments in MLflow

---

What You're Looking At

The Core (Start Here)

complete/fl/
├── fl/
│   ├── task.py          # ⭐ Core: train(), test(), load_data()
│   ├── server_app.py    # ⭐ Server: aggregation logic
│   ├── client_app.py    # ⭐ Client: local training
│   ├── aggregation.py   # FedAvg, FedProx, FedNova
│   ├── dp.py            # Differential privacy
│   └── secure_agg.py    # Secure aggregation
├── config/
│   └── default.yaml     # Configuration
└── tests/
    └── test_*.py        # Test suite (80%+ coverage)

Three files to understand the entire system:

1. task.py - Training loop
2. server_app.py - Server logic
3. client_app.py - Client logic

---

What's Implemented

Algorithms

• ✅ FedAvg (Federated Averaging)
• ✅ FedProx (Proximal term for non-IID)
• ✅ FedNova (Normalized averaging)
• ✅ Scaffold (Variance reduction)

Privacy & Security

• ✅ Differential Privacy (DP-SGD, ε-δ guarantees)
• ✅ Secure Aggregation (encrypted updates)
• ✅ Non-IID data (Dirichlet, label skew)

Production

• ✅ Docker deployment
• ✅ MLflow tracking
• ✅ Real-time dashboard
• ✅ 80%+ test coverage
• ✅ CI/CD pipeline

---

Configuration

Edit complete/fl/config/default.yaml:

topology:
  num_clients: 10
  fraction: 0.5

train:
  lr: 0.01
  local_epochs: 1
  num_server_rounds: 10

data:
  dataset: "albertvillanova/medmnist-v2"
  subset: "pneumoniamnist"
  batch_size: 32

# Optional: Non-IID data
  non_iid:
    type: "label_skew"
    params:
      alpha: 0.5

# Optional: Differential Privacy
privacy:
  dp_sgd:
    enabled: true
    noise_multiplier: 0.8
    target_epsilon: 3.0

---

Performance

Configuration	Dataset	Accuracy	Rounds	Time
FedAvg (IID)	MNIST	98.5%	10	90s
FedAvg (non-IID)	MNIST	96.8%	20	180s
FedProx	MNIST	97.3%	20	195s
FedAvg + DP	MNIST	97.2%	15	120s
FedAvg	PneumoniaMNIST	94.7%	50	300s

Tested on: Intel i7-10700K, 32GB RAM, NVIDIA RTX 3080

---

Development

Local (without Docker)

cd complete/fl
pip install -e ".[dev]"
pytest tests/ -v --cov=fl
flwr run . local-simulation --stream

Docker

./launch-platform.sh  # Start
docker compose -f complete/compose-with-ui.yml down  # Stop

---

Architecture

SuperLink (Server)
    ├── Aggregates updates (FedAvg/FedProx)
    ├── Manages rounds
    └── Logs to MLflow
         │
    ┌────┴────┬────────┬────────┐
    │         │        │        │
SuperNode  SuperNode  ...  SuperNode
(Client 1) (Client 2)     (Client N)
    │         │              │
Private    Private       Private
Data       Data          Data

---

Documentation

• ARCHITECTURE.md - System design
• TROUBLESHOOTING.md - Common issues
• DEPLOYMENT.md - Production deployment
• CONTRIBUTING.md - Development guide

---

Testing

cd complete/fl

# All tests
pytest tests/ -v

# With coverage
pytest tests/ -v --cov=fl --cov-report=html

# Specific test
pytest tests/test_integration.py -v

Coverage: 80%+

---

What Makes This Production-Ready

1. Battle-tested algorithms: FedAvg, FedProx, DP-SGD
2. Comprehensive testing: 80%+ coverage
3. Production deployment: Docker, Kubernetes
4. Privacy guarantees: Differential privacy (ε-δ)
5. Real-world datasets: Medical imaging
6. Experiment tracking: MLflow integration
7. Professional docs: Complete guides

---

License

Apache 2.0

---

One Command to Rule Them All

./launch-platform.sh

Everything else is in the docs.