ResFed-IDS: Resource-Aware Federated Learning for Sustainable IoT Intrusion Detection

Abstract

Federated learning (FL) reduces raw-data sharing in Internet of Things (IoT) intrusion detection systems (IDSs), but standard FL can still overuse battery-powered clients by assigning local training without considering device health. This paper presents ResFed-IDS, a resource-aware FL framework that combines server-side healthy-client selection with client-side self-preservation. Clients are eligible only when battery is at least 30% and central processing unit (CPU) load is at most 0.85; the server then selects up to ⌈0.6K⌉ healthy clients per round and aggregates successful updates through sample-weighted federated averaging (FedAvg). On a balanced CICIoT2023 subset, ResFed-IDS reached 82.11% best accuracy in a 15-round, 5-client simulation with zero device depletion, whereas standard FedAvg reached 81.56% and produced 27 client-unavailability events with four ultimately depleted clients. The same policy transferred to CICIIoT2025 and CIC-ToN-IoT, yielding 78.87% and 69.35% best accuracy while preserving all devices. Relative to the centralized SimpleMLP baseline, the remaining gap is only 0.73 percentage points, indicating that most loss is architectural rather than caused by the resource-aware FL procedure.