Long-Term Energy Management Empowered Hierarchical Federated Learning for Smart Consumer Electronics

Managing for local data among the increasingly popular consumer smart electronics in a secure manner while increasing the user experience is still a hard work. Federated learning (FL), can develop intelligent electronics-related applications while protecting the privacy of local data. However, considering the traditional cloud-based FL training process, consumer electronics with a limited energy budget can reduce the training efficiency. Hence, in this paper, to reduce the total latency of FL training while also meeting a targeted minimum value of loss function and meeting the long-term energy consumption among smart consumer electronics, we introduce an energy-efficient hierarchical FL algorithm and formulate a multi-objective optimization problem including diversified resource allocation and device association. Considering channel state information is unavailable for all rounds, applying the Lyapunov optimization framework, an alternative problem incorporating the significance of local models is reformulated to decrease the training latency per round and enhance long-term performance at the same time. To achieve a better solution to the device association problem, a low-complexity two-operation device association algorithm is proposed, along with resource allocation for training time control, local computing power control, and bandwidth allocation. Our proposed algorithm can achieve better learning performance while meeting the energy budget compared with multiple benchmarks, according to numerical results.