Mini Federated Learning Simulator

Several clients each train a small neural network on their own local data, then a server averages their models into one global model — no raw data ever leaves a client. Tune the scenario below and watch how per-client and global accuracy respond. The training is real (tiny MLP, hand-written backprop), not a mock-up.

Scenario

Hospitals / clients 5

How many participants train separately.

Data imbalance (non-IID) 90

0 = every client sees all classes evenly · 100 = each client is dominated by one class.

Aggregation method

How the server combines the client models each round.

Local epochs 8

Training passes each client does before syncing. High + imbalanced ⇒ drift.

Communication rounds 15

Train → average → repeat, this many times.

seed: — · same settings reproduce the same run

How does this work?

Each round, every client copies the current global model, trains it for a few epochs on its local slice of a synthetic 2-D, 3-class dataset, and sends the weights back. The server averages those weights (FedAvg) into a new global model. Accuracy is measured on one shared held-out test set.

Raising imbalance skews each client toward different classes (via a Dirichlet split), so their models disagree — exactly the hard case federated learning has to cope with.

Even normalization is federated: clients share only count/sum/sum² aggregates of their raw points, and the server combines those into a global mean/std that everyone standardizes with. The held-out test set never touches the stats. (Real deployments would protect even these aggregates with secure aggregation.)

Results

Accuracy over rounds

Data distribution per client

Class 0 Class 1 Class 2 bar height = number of samples

Decision boundaries

Each client's local model (this round) vs. the aggregated global model. Divergent tiles = client drift.

Honest by design: models are genuinely trained in your browser with hand-written gradient descent — no libraries, no network, no faked numbers — and even normalization is federated, built from client aggregates rather than pooled raw data. Data is synthetic 2-D Gaussian blobs so decision boundaries can be drawn. Open index.html directly; it runs fully offline.