Federated Learning for Energy-balanced Client Selection in Mobile Edge Computing

Abstract

Mobile edge computing (MEC) has been consideredas a promising technology to provide seamless integration ofmultiple application services. Federated learning (FL) is carriedout at edge clients in MEC for privacy-preserving training ofdata processing models. Despite that the edge clients with smalldata payloads consume less energy on FL training, the small datapayload gives rise to a low learning accuracy due to insufficientinput to the FL training. Inadequate selection of the edge clientscan result in a large energy consumption at the edge clients,or a low learning accuracy of the FL training. In this paper,a new FL-based client selection optimization is proposed tobalance the trade-off between energy consumption of the edgeclients and the learning accuracy of FL. We first show thatthis optimization problem is NP-complete. Next, we proposea FL-based energy-accuracy balancing heuristic algorithm toapproximate the optimal client selection in polynomial time. Thenumerical results show the advantage of our proposed algorithm.

Exploring Deep Reinforcement Learning-Assisted Federated Learning for Online Resource Allocation in Privacy-Persevering EdgeIoT

“…For performance validation, we compare FL-DLT3 with existing state-of-the-art FL-based device scheduling approaches, i.e., FedAECS [21], FedCS [12] and FedAvg [9].…”

“…An FL-based device selection optimization is developed in our preliminary work [21] to balance the energy consumption of the user devices and the learning accuracy of FL. The optimization model takes advantage of the a-priori knowledge of the network state information, e.g., data size, bandwidth, and channel gain.…”

“…The optimization model takes advantage of the a-priori knowledge of the network state information, e.g., data size, bandwidth, and channel gain. Due to NP-hardness of the optimization, an energy efficiency-FL accuracy balancing heuristic algorithm (FedAECS) was presented in [21] to approximate the optimal client selection policy offline. In contrast, this paper considers a practical scenario without the prior information about data size, bandwidth, channel gain, and remaining energy of the IoT devices.…”

“…Due to a large state and action space, we propose a new FL-DLT3 to balance FL accuracy and energy consumption of the IoT devices online, where LSTM is leveraged to predict the hidden state of the IoT devices as input state of TD3. In addition, we also compare the performance of the proposed FL-DLT3 with the FedAECS in [21].…”

Client Selection and Bandwidth Allocation for Federated Learning: An Online Optimization Perspective

“…Many existing works mainly consider optimizing the FL accuracy within a resource limit in each individual round, however the latency and energy consumption are as important as the accuracy in a real-world federated learning, so how to achieve the appropriate tradeoff between these targets is quite important. On the other hand, many existing works only consider the optimal problem in each individual round omitting the dependence of different rounds [4][5][6], Particularly, early effort on long-term FL optimization is presented in [7]. The author maximizes the linear function which is empirically proportional to the accuracy within the energy limit, but it does not consider the time latency of FL.…”