FLInterrupt: An interactive federated learning simulator for client interruption experiments

Abstract

We introduce FLInterrupt, an open-source, browser-based federated learning (FL) simulator for research on involuntary client interruption during synchronous FedAvg. The backend runs communication rounds in PyTorch. A React dashboard lets researchers set datasets and models, interrupt or reconnect clients at runtime, replay a deterministic JSON interruption schedule, and inspect global metrics, a participation strip, held-out fold evaluation of the frozen global model, and class-wise diagnostics. When a client is interrupted, it does not submit an update for that round; partial local work is discarded. The server aggregates only remaining updates, or leaves the global model unchanged if none arrive. FLInterrupt does not introduce a new aggregator. Client selection and planned exclusion are established FL design patterns. The software contribution is an interactive research loop that treats unexpected unavailability as a controllable factor, with live visualization and exportable experiment JSON, complementary to production FL frameworks.