Project idea · ungraded

Federated Training across Departments

Simulate several college departments that want a shared model (for example, predicting which students need extra support) without pooling their records. Compare federated training with central training, and see how uneven, non-IID data across departments affects the result.

You will use

Milestones

Build the smallest working version first. Ticks are saved in this browser only.

01Generate or load a tabular dataset and split it into 4–6 'departments' with different label mixes.
02Train a central model on all data as the reference.
03Run FedAvg with the PyTorch steps and plot global accuracy per round.
04Vary local epochs and client sampling and record how many rounds each setting needs.
05Add update clipping and noise (differential privacy) and measure the accuracy cost.

Stretch goals

Implement FedProx in PyTorch and compare on the most uneven split.
Run the same experiment with the Flower framework, with clients as separate processes.
Personalise: fine-tune the global model on each department and report per-department accuracy.

Data

Synthetic data from sklearn.datasets.make_classification works well; split it by a feature to create non-IID departments. See the Flower framework documentation for running real federated clients.

Run it yourself

pip install torch scikit-learn flwr
python simulate.py --clients 5 --rounds 30