Trust- and Energy-Aware Federated Learning for Robust and Sustainable IoT Edge Intelligence: Full-System Evaluation on PAMAP2 and Cross-Domain Robustness Assessment on CICIoT2023

Federated learning (FL) supports collaborative edge intelligence without centralizing raw data, but non-IID data, Byzantine attacks, and limited client energy complicate IoT deployment. We study two complementary server-side mechanisms: TEA-FL, which couples consistency-based trust with residual energy in client selection and aggregation, and ARL-FL, which separates risk/lifetime-aware scheduling from pressure-adaptive robust aggregation. A matched two-domain protocol distinguishes full-system PAMAP2 evidence from aggregation-focused CICIoT2023 evidence and combines fold/seed-level inference with ablations, targeted sensitivities, and complete-round overhead measurements. Under the primary simulated resource model, TEA-FL achieved the highest overall PAMAP2 Macro-F1 and lowest normalized energy, but reduced participation fairness and did not complete every nominal-horizon run. ARL-FL preserved the largest active-client population and completed all PAMAP2 runs, but failed under the strongest sign-flip attack. Focused controls showed that trust drives TEA-FL’s sign-flip robustness, energy-aware selection drives ARL-FL’s sustainability benefit, and simpler fixed clipping-plus-trimming can be competitive under strong attack. The results support context-dependent robustness–sustainability trade-offs rather than universal superiority.