Honest gradient Poisoned gradient Global model trail
drag to pan
scroll / pinch to zoom
distance impact %

Byzantine-Robust Federated Learning Defense (2D Vector-Field View)

Federated learning trains one shared model from many clients' local gradient updates without ever centralizing their raw data — but that also means the server must trust updates it cannot inspect. This 2D simulator renders six clients arranged around a central global model: each round every client sends a local update vector, one or more can be set malicious and send a poisoned update scaled away from the shared true optimum, and the server aggregates all six into the next model position. A draggable, zoomable vector-field view shows every gradient arrow live, while a companion strip chart tracks the model's distance from the true optimum and the measured poisoning impact on the aggregate round by round. Switch the aggregation rule between a naive mean — which a single unbounded malicious update can drag anywhere — and a Byzantine-robust coordinate-wise median, which outvotes a minority of poisoned updates instead of averaging them in.