FedCuraLM: Reliability-Calibrated and Resource-Aware Label Auditing with Small Language Models for Non-IID Federated AIoT

Federated Artificial Intelligence of Things (AIoT) activity-recognition systems learn from decentralized wearable and smart-home records with potentially incorrect labels, non-IID client distributions, and unequal edge-device resources. Our objective is to improve downstream learning from these labels while limiting harmful corrections and respecting auditing, review, and communication budgets. FedCuraLM addresses this objective by converting sensor records into deterministic Semantic Evidence Cards, obtaining candidate-constrained small language model (SLM) assessments, and calibrating and combining them with task-model, prototype, and temporal evidence. Conflict-aware decisions retain observed labels, apply provisional soft corrections, request trusted review, or quarantine uncertain samples without overwriting annotations. Review-grounded reliability and communication-budgeted multi-prototype exchange support collaboration, while resource-aware scheduling assigns feasible SLM tiers across heterogeneous devices. Extensive evaluation on WISDM-W, WISDM-P, and CASAS using controlled symmetric, semantic-asymmetric, instance-dependent, and temporal-boundary corruption under varying client heterogeneity reveals that FedCuraLM significantly improves downstream performance over recent baselines and improves lower-tail client outcomes, audit precision, selective-correction safety, calibration, and communication efficiency. Edge-device measurements characterize latency, memory, and energy trade-offs. The findings position federated label curation as a semantic, statistical, communication, and systems problem.