FedRCS-IoT: Availability-Risk-Aware and Fair Client Selection for Federated IoT Intrusion Detection

Federated intrusion detection can keep raw traffic records at the edge, reducing the need to centralize sensitive device data and the exposure associated with central collection. This architectural privacy benefit does not by itself protect model updates or provide a formal privacy guarantee. A training round still depends on clients finishing local work and returning updates before a deadline. Random participation wastes computation when resources fluctuate, whereas selecting only fast clients can repeatedly exclude data-bearing devices. This paper presents FedRCS-IoT, an online client selector that combines calibrated return probability, reference-set utility, a virtual fairness queue, and an exploration bonus under count and airtime budgets. Here “risk” means operational non-return before the round deadline, not malicious-client or privacy risk. The learning rule remains ordinary cross-entropy FedAvg so that scheduling effects are not mixed with an optimization change. We study time-correlated resource states and burst failures on N-BaIoT and Edge-IIoTset. Across ten development seeds, FedRCS improves validation Macro-F1 over Random by 0.0201/0.0141 and reduces simulated completion by 19.1%/14.2% and waste by
57.8%/50.7%, respectively. Against Resource-Only, it raises NBaIoT logical-client Jain fairness from 0.4321 to 0.6833. On a once-only held-out Edge-IIoTset test, FedRCS improves Macro-F1 over Resource-Only by 0.0261 (paired 95% CI [0.0093, 0.0429]); frozen held-out differences from Random remain unresolved. The evidence supports a conditional utility–efficiency–fairness tradeoff, not universal superiority.