AF-IoTS: An Adaptive Trust-Aware Federated Architecture for Scalable Intrusion Detection in Heterogeneous IoT Environments

Abstract

The Internet of Things (IoT) generates massive, privacy-sensitive traffic across heterogeneous, resource-constrained devices, making centralized intrusion detection systems (IDS) increasingly impractical due to scalability, latency, and privacy limitations. Existing federated learning (FL) based IDS solutions partially mitigate these issues but typically assume static training schedules, uniform participation, and reliable clients, and therefore remain vulnerable to non-IID data distributions, device heterogeneity, and adversarial behaviors such as poisoning and Byzantine attacks. This paper proposes Adaptive Federated IoT Security (AF-IoTS). This architectural framework integrates federated learning with three coordinated mechanisms: (i) a trust model that combines behavioral deviation, loss, uncertainty, availability, and resource ratio into an exponential moving-average trust score; (ii) a diversity-aware client-selection utility that jointly considers Trust, resources, and data diversity under communication and cardinality budgets; and (iii) a trust-weighted robust aggregation rule with update clipping, driven by a closed-loop adaptive controller that regulates the participation rate. The proposed workflow is formalized through explicit mathematical formulations and an end-to-end algorithm, and its behavior is analyzed qualitatively across smart cities, smart healthcare, and industrial IoT scenarios. Expected outcomes include preserving data locality, improving robustness against unreliable and malicious clients, and providing a coherent architectural basis for future large-scale empirical validation on standard IoT security datasets.