Confidence-calibrated federated graph attention for internet of things agents under latency SLOs

Internet of Things (IoT) agents that trigger network enforcement actions must be both well-calibrated (for safe triage) and tail-latency predictable (for service level objectives, SLOs). We present Confidence-Calibrated HP-FedGAT-Trust-IBN, a federated, graph-attention architecture that closes the loop from IoMT sensing to SDN enforcement via parameter-efficient (LoRA/PEFT) updates ([Formula: see text] MB/round), trust-weighted secure aggregation, and intent verification (IBN) triage. Evaluation follows a two-plane protocol: a learning plane with [Formula: see text] simulated clients under a matched comparator harness (Graph-FL and uncertainty-aware FL baselines), and a serving plane that replays exported checkpoints on real edge devices (Raspberry Pi 5, Jetson Orin Nano, Intel NUC 11) and validates SLOs using hardware ECDFs and empirical [Formula: see text]. The model achieves high discrimination (ROC-AUC/PR-AUC [Formula: see text]-[Formula: see text]) with improved calibration (low ECE) under the matched harness, while the serving loop satisfies the [Formula: see text] ms requirement by device-measured [Formula: see text] (e.g., enforcement [Formula: see text] ms, vs. [Formula: see text] ms for an efficient-UQ baseline) and explicit compliance [Formula: see text]. The latency decomposition includes all calibration costs and Monte-Carlo expectations ([Formula: see text], with measured MC share reported), and security modes are quantified end-to-end: CKKS + SMPC adds device-measured [Formula: see text] and crypto-attributable Joules (e.g., [Formula: see text] ms and [Formula: see text] J/round on Raspberry Pi 5). Energy/round is measured on identical hardware and mapped to CO₂^e for carbon-aware selection of operating points.