Energy- and QoS-Gated Federated Learning for Sustainable Wearable Stress Detection

Federated Learning (FL) is increasingly deployed on battery-powered edge and wearable devices, where every communication round and local training step carries a direct energy and carbon cost. Most FL evaluations, however, still report predictive accuracy alone, leaving the environmental footprint of the learning process itself unmeasured. This paper presents a two-phase, QoS- and energy-gated FL approach, FedEQ, which restricts local training to moments when a client's sensing buffer and residual energy meet safety thresholds, and restricts transmission to updates whose novelty justifies their instantaneous latency and energy cost. We evaluate FedEQ on a federated stress-detection task built from the public WESAD dataset, benchmarking it against three established aggregation schemes: FedAvg, FedProx, and the adaptive server-side optimizer FedAdam. Beyond accuracy and F1, we report communication volume, wall-clock-measured compute energy (via CodeCarbon), analytically modeled communication energy, CO2eq emissions, and floating-point operations, averaged over five random seeds. FedAdam attains the highest raw accuracy of the four methods (93.2%) at the shortest horizon we test (20 rounds); FedEQ retains 97.0% of that accuracy at the same horizon while cutting communication volume by 63.9%, total energy and CO2eq by 68.9-69.7%, and FLOPs by 78.6% relative to all three baselines, for 2.7-3.2x higher accuracy-per-resource-unit (per MB, per Wh, and per gCO2eq) than the strongest competing method. We further show this accuracy gap is a short-horizon artifact rather than a fixed property of the method: extending training to 60 rounds closes it to within 0.5 points of the best baseline - and FedEQ overtakes FedAdam outright - while its communication savings simultaneously grow to 69.1%. Across the full range tested, FedEQ is the Pareto-dominant choice among the four once predictive performance and resource cost are considered jointly.