Research modern TensorFlow Federated implementation

[iAmGiG]

Background

The 2020 research attempted to implement federated learning using TensorFlow Federated (TFF) 0.40.0, but:

1. TFF 0.40.0 has unresolvable dependency conflicts (jaxlib==0.3.14 doesn't exist for Python 3.9+)
2. The FL code (train_federated_simulation.py) never successfully used botnet data - it still uses EMNIST digits
3. Time pressure + API complexity led to incomplete adaptation

See: docs/project-analysis/RETROSPECTIVE.md lines 103-143

Goals for Modern Branch

1. Evaluate modern TFF versions:

   • TFF 0.60+ (2024 releases)
   • Check if API has improved since 2020
   • Verify TensorFlow 2.15+ compatibility

2. Alternative frameworks:

   • Flower - Modern FL framework, framework-agnostic
   • Better documentation than 2020-era TFF
   • Active community support

3. Implementation research:

   • Can we adapt tabular data (115 features) to FL architecture?
   • Simulate multiple IoT devices as FL clients
   • Compare centralized vs federated training results

4. Stretch goal:

   • Actually implement FL with N-BaIoT dataset
   • Per-device models with federated aggregation
   • Demonstrate privacy-preserving botnet detection

Success Criteria

• Survey modern FL frameworks (TFF, Flower, PySyft)
• Prototype FL implementation with botnet data
• Compare accuracy: centralized vs federated
• Document findings and recommendations

Notes

• This is research/exploration, not a blocker
• Focus on modern main branch, not archive branches
• Original 2020 attempt was valid for the time - tools have improved
• May result in new paper/publication if successful

[iAmGiG]

Modernization plan created. Will replace TensorFlow Federated with Flower framework. See docs/MODERNIZATION_ROADMAP.md Phase 5 and issue #28 for implementation details.