Implement Federated Learning with Flower

[iAmGiG]

Goal

Replace incomplete TensorFlow Federated implementation with modern Flower framework.

Background

Current State (TFF):

• TensorFlow Federated 0.40.0
• Impossible to install (dependency conflicts)
• Never successfully used botnet data (used EMNIST instead)
• Simulation-only, not production-ready

Target (Flower):

• Modern, actively maintained
• Better documentation
• Framework-agnostic
• Production deployment support

Implementation

Architecture

src/federated_learning/
├── __init__.py
├── flower/
│   ├── __init__.py
│   ├── server.py          # Flower server
│   ├── client.py          # Device clients
│   └── strategy.py        # Custom FedAvg
├── simulation/
│   ├── __init__.py
│   └── run_simulation.py  # FL simulation
└── utils/
    ├── __init__.py
    └── aggregation.py

Components

Server (server.py):

• Orchest rates federated training
• Aggregates client updates
• Custom FedAvg strategy

Client (client.py):

• One client per IoT device (9 total)
• Local training on device data
• Send updates to server

Simulation (run_simulation.py):

• Simulated FL environment
• Test before deployment

Tasks

• Install Flower (flwr)
• Implement server
• Implement client
• Implement custom aggregation strategy
• Create simulation framework
• Run experiments
• Compare centralized vs federated accuracy
• Document FL implementation
• Create deployment guide

Experiments

1. Baseline: Centralized training (all devices)
2. Local: Per-device models (no aggregation)
3. Federated: Flower FL (FedAvg)

Metrics:

• Accuracy (global and per-device)
• Communication rounds
• Training time
• Model size

Research Questions

1. Does FL maintain accuracy vs centralized?
2. Can we achieve device-specific personalization?
3. What is the communication cost?
4. How does it scale?

Related

• Replaces Fix deprecated pandas.append() calls #14 (old TFF issue)
• Part of MODERNIZATION_ROADMAP.md Phase 5.1
• Related to Research modern TensorFlow Federated implementation #22 (FL research)

Priority

HIGH - Core feature for federated learning capability

[iAmGiG]

Good news: TensorFlow 2.19.1 is available (even newer than planned 2.18). Environment updated to use Python 3.12.12 + TensorFlow 2.19.1 + Keras 3.6+ for maximum performance.