JS-Drift: A reproducible Jensen-Shannon divergence procedure for drift-aware client weighting in federated learning

Federated learning lets institutions train a shared model without exchanging raw data, but standard aggregation (FedAvg) assumes client data distributions are stationary. In practice they drift over time, and aggregation that ignores this lets unstable clients degrade the global model. This article describes JS-Drift, a reproducible, model-agnostic procedure that quantifies per-client temporal drift using Jensen-Shannon (JS) divergence between a client's label distribution in consecutive communication rounds, converts that divergence into a stability weight via a single sensitivity parameter γ, and folds the weight into the aggregation step. The procedure requires no architectural changes, adds negligible overhead, and drops into any FedAvg-style training loop. We give the full algorithm, exact computation steps, parameter-selection guidance, and an open implementation, and we validate that the procedure behaves as intended on three structurally different tabular-classification settings.•Computes a per-client, per-round drift coefficient from JS divergence between consecutive local label distributions; only a small class-proportion summary is shared, so raw data never leave the client.•Maps the drift coefficient to an aggregation weight through one interpretable parameter γ, down-weighting clients with high distributional shift.•Is model-agnostic and integrates into any FedAvg-style round in a few lines of code; a public repository reproduces every step.