BEFL: Balancing Energy Consumption in Federated Learning for Mobile Edge IoT
Zehao Ju 1 , Fuke Shen 2 , Tongquan Wei 3,* · 2024
Read BEFL: Balancing Energy Consumption in Federated Learning for Mobile Edge IoT by Zehao Ju 1 , Fuke Shen 2 , Tongquan Wei 3,* (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-03950.txt.
Introduction
Zehao Ju and co-authors tackle battery drain in federated learning for mobile edge devices in a 2024 paper. In this setup, many small devices train a shared model without sending raw data. Training and messaging can drain weak batteries, and cutting total use alone can leave some devices to drop out early.
Their BEFL frame balances three goals: model accuracy, total energy, and fairness of energy use across devices. It allocates communication resources with a math solver and picks clients with a clustering heuristic. The aim is a model that stays strong without burning out the weakest members of the fleet.
Catalog record
| CREATOR | Zehao Ju 1 , Fuke Shen 2 , Tongquan Wei 3,* |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 20 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-03950
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
BEFL: Balancing Energy Consumption in Federated Learning for Mobile Edge IoT Zehao Ju 1 , Fuke Shen 2 , Tongquan Wei 3,* Abstract Federated Learning (FL) is a privacy-preserving distributed learning paradigm designed to build a highly accurate global model. In Mobile Edge IoT (MEIoT), the training and communication processes can significantly deplete the limited battery resources of devices. Existing research primarily focuses on reducing overall energy consumption, but this may inadvertently create energy consumption imbalances, leading to the premature dropout of energy-sensitive devices.To address these challenges, we propose BEFL, a joint optimization framework aimed at balancing three objectives: enhancing global model accuracy, minimizing total energy consumption, and reducing energy usage disparities among devices. First, taking into account the communication constraints of MEIoT and the heterogeneity of devices, we employed the Sequential Least Squares Programming (SLSQP) algorithm for the rational allocation of communication resources. Based on this, we introduce a heuristic client selection algorithm that combines cluster partitioning with utility-driven approaches to alleviate both the total energy consumption of all devices and the discrepancies in energy usage.Furthermore, we utilize the proposed heuristic client selection algorithm as a template for offline imitation learning during pre-training, while adopting a ranking-based reinforcement learning approach online to further boost training efficiency. Our experiments reveal that BEFL improves global model accuracy by 1.6%, reduces energy consumption variance by 72.7%, and lowers total energy consumption by 28.2% compared to existing methods. The relevant code can be found at https://github.com/juzehao/BEFL . I Introduction Federated Learning (FL) is a distributed and iterative machine learning approach that builds a global model while preserving user privacy by not requiring access to raw data [ 2 , 3 ] . However, deploying FL in Mobile Edge IoT (MEIoT) environments introduces significant energy consumption challenges. Many IoT devices are battery-powered [ 4 ] , which restricts their energy capacity. Consequently, energy optimization in FL has become a focal point for researchers. For example, Kang et al. [ 6 ] proposed the ”Neurosurgeon” scheduler, which offloads parts of deep neural network (DNN) computations from mobile devices to data centers, reducing both latency and energy consumption in industrial IoT networks. Zhang et al. [ 7 ] applied techniques like filter pruning and multitask learning to compress DNN models, lowering energy usage. However, these approaches often fail to consider the device heterogeneity present in MEIoT environments, which significantly affects energy consumption optimization. Fig. 1 : The two issues in energy consumption optimization for FL: 1) The impact of device heterogeneity on energy consumption ensitivity. 2) The imbalance in energy consumption among devices. In FL, leveraging heterogeneous devices is essential to improve data diversity and model accuracy. This heterogeneity leads to varying energy and time demands for tasks like model training and data communication. Ignoring these differences can result in excessive reliance on energy-hungry devices, increasing overall energy consumption and causing workload imbalances. To mitigate these issues, Cui et al. [ 9 ] designed a scheduling strategy that uses frequency control techniques to reduce energy costs while maintaining model accuracy and performance across heterogeneous devices. Moreover, Tian et al. [ 17 ] introduced FedRank, a novel client selection approach using a ranking-based mechanism trained with imitation learning and reinforcement learning (RL) which optimizes the selection process by considering data and system heterogeneity. Although these works primarily address the issue of heightened overall energy consumption due to device heterogeneity, they fall short in tackling energy imbalance and the differing energy sensitivities among devices. This oversight may result in excessive energy use for certain devices or their premature withdrawal, ultimately hindering the progress of FL. In light of this, we present the following contributions: • To address client energy consumption imbalance and excessive use of energy-sensitive devices in the MEIoT environment, we recommend leveraging relative energy consumption and incorporating energy balance into the system’s optimization objectives. • Given device heterogeneity and communication resource constraints, we apply the SLSQP algorithm for resource allocation and introduce a heuristic client selection algorithm combining clustering and utility-driven methods to enhance federated learning efficiency and reduce energy imbalance. • For optimal client selection during model training, we employ the BEFL framework, which uses offline heuristic learning and online RL. Experiments show that BEFL synchronizes model performance, total client energy consumption, and energy variance in dynamic training. II SYSTEM MODEL AND PROBLEM FORMULATION II-A System Model We consider a MEIoT scenario with an edge server and multiple client devices. In each FL round, the server sends the global model to selected clients \phi_{h}=\{\theta_{1},\ldots,\theta_{n}\} . Each client \theta_{i}\in\phi_{h} trains the model on its local dataset, incurring energy E_{i}^{train} and latency \tau_{i}^{train} . After training, the client uploads the model parameters, resulting in additional energy E_{i}^{trans} and latency \tau_{i}^{trans} . II-A 1 Local training model The training time for device \theta_{i} is defined as \tau_{i}^{train}=\frac{I_{i}C_{i}|D_{i}|}{f_{i}}core_{i} , where C_{i} , I_{i} , |D_{i}| , f_{i} , and core_{i} represent CPU cycles per sample, local iterations, number of training samples, CPU frequency, and cores, respectively. The training energy consumption is given by E_{i}^{train}=\kappa I_{i}C_{i}|D_{i}|f_{i}^{2}core_{i} with \kappa being the capacitance coefficient of the device. II-A 2 Local communication model The communication method using Orthogonal Frequency Division Multiple Access (OFDMA), the communication rate for device is \theta_{i} is r^{trans}_{i}=\beta_{i}B\log_{2}\left(1+\frac{g_{i}^{2}P_{i}}{N_{0}}\right) , where \beta_{i} is the resource allocation ratio, B is bandwidth, g_{i} is channel gain, P_{i} is communication power, and N_{0} is noise power spectral density. The communication energy consumption is E_{i}^{trans}=P_{i}\frac{G}{r^{trans}_{i}}=P_{i}\frac{G}{\beta_{i}B\log_{2}\left(1+\frac{g_{i}^{2}P_{i}}{N_{0}}\right)} . II-A 3 Relative energy consumption model Total energy consumption includes both training and communication. We define relative energy consumption as E_{i}^{r}=\frac{E_{i}^{train}+E_{i}^{trans}}{E_{i}^{total}\delta} , where E_{i}^{total} is the device’s battery capacity and \delta is the sensitivity coefficient.The energy consumption mentioned in this article specifically refers to the relative energy consumption under this formula. II-B Problem Definition We address a problem within a MEIoT environment comprising an edge server and a set of client devices \Phi=\{\theta_{1},\theta_{2},\ldots,\theta_{n}\} . The energy consumption of each client during FL training is represented by \Xi=\{E_{1},E_{2},\ldots,E_{n}\} , across a total of H training rounds. The set of devices selected in the h -th round is denoted as \phi_{h} . Our objectives are to minimize the variance in energy consumption among clients, reduce the overall energy usage, and maximize the global model’s accuracy. Objective: \min\{\text{Var}(\Xi),\max_{\theta_{i}\in\Phi}E_{i},\sum_{h=1}^{H}E_{\phi_{h}},L(w_{g},D_{all})\} . Constraints: \displaystyle E_{\phi_{h}} \displaystyle\geq\sum_{\theta_{i}\in\phi_{h}}(E_{i}^{train}+E_{i}^{trans}), (1) \displaystyle T_{\phi_{h}} \displaystyle\geq\max_{\theta_{i}\in\phi_{h}}(\tau_{i}^{train}+\tau_{i}^{trans}), (2) \displaystyle\sum_{\theta_{i}\in\phi_{h}}\beta_{i} \displaystyle\leq 1, (3) \displaystyle\beta_{i} \displaystyle\in[0,1],\quad\theta_{i}\in\phi_{h}, (4) \displaystyle\sum_{h=1}^{H}T_{\phi_{h}} \displaystyle\leq T_{limit}. (5) These constraints ensure adequate energy allocation for training, adherence to synchronization limitations, and proper distribution of total resources within predefined limits. Due to the complexity of the issue, we will investigate heuristic scheduling methods to manage the NP-hard characteristics of the problem. III Framework Design In this section, we present the framework’s design, including the communication resource allocation strategy, the heuristic method based on offline imitation learning, and definitions related to online RL. Fig. 2 : BEFL framework overview: 1. The server retrieves hardware information from devices. 2. It simulates energy consumption and latency using energy consumption model. 3. Client selection is facilitated by a heuristic algorithm. 4. Generated state-action pairs are sent to the RL network for pre-training. 5-11. During online training, the RL agent gathers state information, selects clients, conducts local training, aggregates parameters, and updates both main and target