ARotting/federated-digits

Federated Digits

Federated Digits simulates five data-siloed clients learning a shared 8x8 digit classifier with Federated Averaging. Class samples are partitioned through a low-concentration Dirichlet distribution, producing deliberately non-IID clients.

The project reports:

• per-client class distributions;
• global test accuracy after every communication round;
• client-update disagreement;
• a centralized training control using the same architecture.

No raw client examples are exchanged during federated rounds; only model parameters are averaged. This is a local simulation, not a networked privacy guarantee.

Reproduce

uv run python projects/federated-digits/train.py

Verified results

• Model size: 2,410 parameters
• Clients: 5, with Dirichlet concentration 0.25
• Communication rounds: 30, with two local epochs per round
• Round-one global accuracy: 20.89%
• Final global accuracy: 93.33%
• Best global accuracy: 94.00%
• Centralized control: 96.89%
• Final federated-to-centralized gap: -3.56 percentage points

Mean client-update disagreement fell from 1.487 after the first round to 0.500 after round 30. The artifact includes every round's accuracy, client loss, update disagreement, and each client's class distribution.