Distributed edge systems require federated learning (FL) frameworks that can jointly support timely model updates, user-level personalization, and communication-efficient coordination. However, conventional hierarchical federated learning (HFL) methods mainly focus on reducing communication overhead and often struggle to maintain model freshness and local adaptability in multi-tier environments with heterogeneous data and resource constraints. To address these challenges, we propose an Energy- and Age-of-Information (AoI)-Aware Hierarchical Personalized Federated Learning (EA-HPFL) framework. The proposed method introduces two key ideas: 1) selective module exchange, in which clients communicate only lightweight adapter, head, or other trainable sub-module updates instead of full model parameters; and 2) deterministic AoI- and energy-aware aggregation, in which freshness and energy indicators guide client- and edge-level aggregation. This design enables efficient hierarchical coordination while preserving local personalization under non-IID data distributions. Experimental results on WESAD, HAR, and Amazon-EN show that EA-HPFL provides a favorable trade-off among communication efficiency, update freshness, and personalized performance across diverse edge settings. In particular, the proposed framework maintains competitive accuracy while substantially reducing the overhead associated with full-model exchange, and the revised TinyLlama-based Amazon-EN evaluation further confirms its effectiveness in an LLM-oriented setting. These results demonstrate the practical value of AoI-aware coordination and selective sub-module exchange for scalable and adaptive FL in real-time intelligent edge systems.

In 5G networks, the Legacy Paging mechanism for device connections faces challenges due to limited resources and increasing demands from human-type communication (HTC) and massive machine-type communications (mMTCs). This results in higher connection setup time (CST) under high traffic. Group paging, developed for mMTC, allows multiple devices to respond to a single paging message, but increases energy consumption for HTC due to irregular data transmission patterns. To address this, we propose the deep reinforcement learning-based group configuration control (DRL-GCC) mechanism. DRL-GCC combines legacy and group paging methods to optimize paging for HTC by dynamically adjusting group configurations, using the proximal policy optimization algorithm to reduce energy consumption and satisfy the required CST. Simulation results demonstrate that DRL-GCC significantly reduces energy consumption in the random access channel (RACH) procedure and the overall energy usage of devices while ensuring compliance with the CST constraint compared to its alternative methods. Notably, DRL-GCC achieves up to a 43% reduction in energy used during the RACH procedure and a 9% decrease in total energy consumption compared to the existing benchmarks. These improvements imply DRL-GCC’s ability to balance energy efficiency with operational performance, providing a robust solution for efficient HTC-focused service provisioning in 5G networks.

Hybrid Language Models (HLMs) are inference-time architectures that combine the low-latency efficiency of Small Language Models (SLMs) on clients (edge devices) with the high accuracy of Large Language Models (LLMs) in centralized servers. Unlike traditional end-to-end LLM inference, HLMs aim to reduce latency and communication by selectively invoking LLMs only when the local SLM’s predictions are uncertain—that is, when the model exhibits low confidence or high entropy in its token-level probability distribution. However, when the SLM encounters ambiguous or low-confidence predictions during inference, it must offload token-level probability distributions to the LLM for refinement. This frequent offloading leads to substantial communication overhead, particularly in bandwidth-constrained environments. To address this challenge, we propose FedHLM, a communication-efficient HLM framework that integrates uncertainty-aware inference with Federated Learning (FL). The key innovation lies in collaboratively learning token-level uncertainty thresholds that determine when SLM predictions require LLM assistance. Instead of relying on static or hand-tuned thresholds, FedHLM uses FL to enable distributed threshold optimization across clients while preserving data privacy. Additionally, embedding-based token representations are employed to facilitate semantic similarity comparisons during Peer-to-Peer (P2P) resolution, allowing clients to reuse tokens inferred by similar peers without efficiently involving the LLM. Moreover, we propose hierarchical model aggregation as a strategy to reduce redundant token transmissions. At the edge server level, client updates are aggregated to refine local routing policies, while global coordination across clusters further synchronizes decision boundaries. This layered approach ensures that repeated uncertainty patterns are captured and resolved locally, significantly reducing unnecessary LLM queries. Extensive simulations on large-scale news classification tasks demonstrate that FedHLM achieves over 95% reduction in LLM transmissions with negligible accuracy loss, highlighting its potential for scalable and efficient edge-Artificial Intelligence (AI) deployment.

In a variety of Internet of Things (IoT) applications, there is a growing need to constantly transmit large amounts of real-time data from IoT sensors to enable precise environmental monitoring. However, this constant transmission of data can result in high energy consumption of IoT sensors, which presents a significant challenge for IoT systems. Therefore, this article presents a new approach for controlling the transmission period of IoT sensors called imputation error cluster-based transmission period control (IEC-TPC) framework, with the goal of reducing energy consumption while maintaining accurate data collection. In order to effectively balance energy consumption and data collection accuracy, the proposed approach uses an imputation error centroid (IEC) model that forms clusters based on imputation error vectors and approximates the centroid of each cluster using a logistic function. Additionally, a new imputation error cluster prediction (IECP) model is designed using long short-term memory (LSTM)-based encoding and convolutional neural network (CNN) models to predict the label of each cluster in advance. By combining the IEC model with a numerically modeled energy consumption function, a min–max optimization problem is properly formulated to minimize the worst-case performance of transmission period control. To find the optimal solution, a Uniqueness Condition function is properly defined, and an optimal objective value searching algorithm (O2VSA) leveraging a bisection search is proposed. Performance evaluation shows that the proposed method achieves similar data collection accuracy in CO2 and humidity data sets while reducing energy consumption by 12.92% and 35.83% than other baseline models, respectively. Moreover, the proposed algorithm outperforms the state-of-the-art method in the temperature data set, with 13% lower energy consumption value. The proposed method also achieves a significant reduction in energy consumption of 95.21% compared to the uniform transmission period model in this data set, while maintaining a low data reconstruction error of only 0.42%. Overall, the proposed method offers a promising approach for efficient and accurate transmission period control in IoT applications.

To realize intelligent network management for future 6G-mobile edge computing (MEC) systems, mobile traffic prediction is crucial. Most of the previous machine learning-driven prediction approaches adopt traditional centralized training paradigm wherein mobile traffic data should be transferred to a central server. To exploit the distributed and parallel processing nature of MEC servers for training mobile traffic prediction models in a fast and secure manner, we propose a novel federated learning (FL) framework wherein locally trained prediction models over MEC servers are aggregated into a global model with joint optimization of MEC server selection and data set management for FL participation. From mathematical investigations of the influence of MEC server participation and data set utilization on the global model accuracy and training costs, including both training latency and energy consumption in the FL process, we first formulate an optimization problem for balancing the accuracy-cost tradeoff by considering a linear accuracy estimation model. Here, the optimization problem is designed using mixed-integer nonlinear programming, which is generally known as NP-hard. We then leverage a number of relaxation techniques to develop near-optimal yet the plausible algorithm based on linear programming. Furthermore, for practical concern, the proposed problem is extended by considering a concave accuracy estimation model; a genetic-based heuristic approach to the extension is proposed for determining the suboptimal solution. The numerical and simulation results suggest that our proposed framework can be effective for building mobile traffic prediction models in a more cost-efficient manner while maintaining competitive prediction accuracy.

This article proposes a novel deep learning-based robust Internet of Things (IoT) sensor data transmission period control (DL-RDTPC) framework in an IoT edge computing system. In general, as the data transmission period of IoT sensors increases, the energy consumption of IoT sensors is reduced, and contrarily, the amount of un-transmitted data (i.e., missing values) becomes continuously accumulated. Therefore, the IoT server is in charge of accurately imputing these missing data for reliable data analysis. By addressing this issue, we newly design the imputation accuracy prediction (IAP) module, which captures the complicated relationships between the imputation accuracy and the data transmission period, in order to estimate the imputation accuracy, precisely. For constructing the IAP, three submodules, which include a stacked bidirectional long short-term memory (Bi-LSTM) model, a multihead convolutional neural network (CNN), and a neural network-based period information encoding network (PIEN) are leveraged. To balance the tradeoff between the imputation accuracy and energy consumption regarding the data transmission period, the multiobjective optimization problem is formulated for minimizing the maximum value of both: 1) the energy consumption of IoT sensors obtained from the analytical model and 2) the imputation accuracy predicted from IAP module. The optimal solution is consequently obtained by utilizing the bisection search algorithm. Extensive performance evaluations validate the effectiveness of the proposed RDTPC algorithm in terms of both the average energy consumption (maximum 68% reduction) and missing data imputation accuracy (maximum 64% RMSE reduction) over other benchmarks. Finally, this article provides a practical implementation of the proposed RDTPC framework via the HTTP protocol under the IEEE 802.11-based WLAN network, as well as interworking with the commercial cloud server.

In order to reduce unnecessary data transmissions from Internet of Things (IoT) sensors, this article proposes a multivariate-time-series-prediction-based adaptive data transmission period control (PBATPC) algorithm for IoT networks. Based on the spatio-temporal correlation between multivariate time-series data, we developed a novel multivariate time-series data encoding scheme utilizing the proposed time-series distance measure$\textit {ADMWD}$. Composed of two significant factors for a multivariate time-series prediction, i.e., the absolute deviation from the mean (ADM) and the weighted differential (WD) distance, the$\textit {ADMWD}$considers both the time distance from a prediction point and a negative correlation between the time-series data concurrently. Utilizing the convolutional neural network (CNN) model, a subset of IoT sensor readings can be predicted from encoded multivariate time-series measurements, and we compared the predicted sensor values with actual readings to obtain the adaptive data transmission period. Extensive performance evaluations show a substantial performance gain of the proposed algorithm in terms of the average power reduction ratio (approximately 12%) and average data reconstruction error (approximately 8.32% MAPE). Finally, this article also provides a practical implementation of the proposed PBATPC algorithm via the HTTP protocol under the IEEE 802.11-based WLAN network.

This article studies distributed computing (DC) mechanisms on heterogeneous mobile devices (MDs) for latency reduction (LR) in Internet-of-Things (IoT) services by mitigating the effect of straggling MDs. We propose novel coded DC mechanisms with two different incentive distributions that consider the time-discounting value of processed results and the amount of the workload computed by MDs. Specifically, we consider distributed gradient descent computing with coding when a task publisher (TP) with a limited amount of budget offers incentives to encourage MDs' participation in the computation. To analyze a hierarchical decision-making structure of the TP and MDs, we formulate a strategic competition between them as a Stackelberg game. In the case that the MDs are the leaders, we design a CPU-cycle frequency control scheme to balance each MD's computing speed and energy consumption for obtaining its maximum utility with the incentive mechanisms. As the follower, the TP aims at minimizing latency of the DC, and it follows the MDs' decisions to determine the load allocation for each MD. Then, we design an algorithm achieving the Stackelberg equilibrium, which is shown to be a unique Nash equilibrium of the game. The performance evaluation results show that the proposed mechanisms achieve 39% of LR on average compared to the benchmark mechanism. Furthermore, the results corroborate the efficiency of the proposed mechanisms in terms of the MDs' social welfare.

This article proposes a novel fair and scalable relay control (FSRC) scheme for the Internet-of-Things (IoT) services in long range (LoRa)-based low-power wide-area networks. The proposed FSRC scheme promotes relay operation with low spreading factor (SF) to improve the success probability for distant end-devices (EDs) and the fairness of success probability for each SF region. To achieve this, a theoretical framework for designing the relay operation is analytically developed by considering a practical LoRaWAN MAC protocol as an analytical model. The proposed FSRC scheme encompasses a selective relay operation by considering both signal-to-noise ratio and signal-to-interference ratio and the receive signal strength indicator value for the location-unaware relay selection strategy. Using this model, a genetic algorithm-based relay control strategy is proposed to maximize both coverage probability and minimum success probability for all SF regions by controlling the relay parameters, such as source-relay region and source-relay ratio. Our numerical analysis validates the effectiveness of the proposed FSRC scheme under various parameters in terms of the minimum success probability of each SF region, coverage probability, and fairness. Specifically, we verify that the proposed FSRC scheme achieves a maximum of approximately 37% and 33% improvement of the minimum success probability and coverage probability, respectively, under practical LoRa PHY/MAC parameters, compared to the single-hop environment (without relay operation).

This paper designs and implements a novel Temporal difference based adaptive object Detection (ToDo) platform in the video analytics at edge computing system. The proposed ToDo platform contains object tracking function and monitoring function, respectively. Based on temporal difference, the proposed ToDo skips the frames for object detection through adaptively controlling object detection rates. Then, it provides a monitoring function of object tracking and resource usages. The proposed ToDo platform is implemented on commercial edge node Jetson TX2 by utilizing YOLO (You Only Look Once) v3, and Dashboard, respectively. We evaluate its performance through extensive measurement-based analysis, and reveal that the proposed ToDo reduces GPU memory footprint to 17% while conducting moving objects detection with 7% accuracy loss.

Voice activity detection (VAD) is a challenging task in low signal-to-noise ratio (SNR) environment, especially in non-stationary noise. To deal with this issue, we propose a novel attention module that can be integrated in Long Short-Term Memory (LSTM). Our proposed attention module refines each LSTM layer's hidden states so as to make it possible to adaptively focus on both time and frequency domain. Experiments are conducted on various noisy conditions using Aurora 4 database. Our proposed method obtains the 95.58 % area under the ROC curve (AUC), achieving 22.05 % relative improvement compared to baseline, with only 2.44 % increase in the number of parameters. Besides, we utilize focal loss for alleviating the performance degradation caused by imbalance between speech and non-speech sections in training sets. The results show that the focal loss can improve the performance in various imbalance situations compared to the cross entropy loss, a commonly used loss function in VAD.

With the widespread use of Internet of Things (IoT), edge computing has recently emerged as a promising technology to tackle low-latency and security issues with personal IoT data. In this regard, many works have been concerned with computing resource allocation of the edge computing server, and some studies have conducted to the pricing schemes for resource allocation additionally. However, few works have attempted to address the comparison among various kinds of pricing schemes. In addition, some schemes have their limitations such as fairness issues on differentiated pricing schemes. To tackle these limitations, this article considered three dynamic pricing mechanisms for resource allocation of edge computing for the IoT environment with a comparative analysis: BID-proportional allocation mechanism (BID-PRAM), uniform pricing mechanism (UNI-PRIM), and fairness-seeking differentiated pricing mechanism (FAID-PRIM). BID-PRAM is newly proposed to overcome the limitation of the auction-based pricing scheme; UNI-PRIM is a basic uniform pricing scheme; FAID-PRIM is newly proposed to tackle the fairness issues of the differentiated pricing scheme. BID-PRAM is formulated as a noncooperative game. UNI-PIM and FAID-PRIM are formulated as a single-leader-multiple-followers Stackelberg game. In each mechanism, the Nash equilibrium (NE) or Stackelberg equilibrium (SE) solution is given with the proof of existence and uniqueness. Numerical results validate the proposed theorems and present a comparative analysis of three mechanisms. Through these analyses, the advantages and disadvantages of each model are identified, to give edge computing service providers guidance on various kinds of pricing schemes.

In this article, to meet a delay requirement for data analytics from Internet-of-Things (IoT) devices, we design a novel market model of the distributed learning resource management mechanism for multiple mobile-edge computing (MEC) operators. We consider a hybrid architecture (cloud-MEC) for the distributed learning, which is also known as “federated learning” as one of the practical examples, where a coordinator at the cloud coordinates IoT sensors to efficiently distribute their sensing data over multiple MECs. In this sense, multiple MECs receive a shared model from the coordinator and conduct local training of received partial sensing data from IoT sensors. Then, the coordinator at the cloud merges returned local training results from MECs and generates a global model. To model a hierarchical decision-making structure as a market behavior, we formulate and solve a Stackelberg game model. Specifically, in the case of MEC operators as leaders, we design a pricing scheme for MEC operators to obtain its maximum utility by considering a tradeoff between the revenue and energy consumption. Then, as a follower, while the coordinator aims at achieving a balance between the satisfaction attained from the distributed learning and the costs, it follows the MEC operators' decisions by coordinating IoT sensors to distribute their sensing data over MECs. A unique Stackelberg equilibrium (SE) point is given as a closed form. Finally, we reveal that the SE solution maximizes the utility of all market participants. This game-theoretic study demonstrates that there is an incentive for utilizing multiple MECs to achieve better satisfaction of IoT services in distributed learning.

This paper proposes auction-based energy trading among electric vehicles (EVs) with consideration of practical battery status. At an arbitrary time, each EV can be a seller or a buyer according to their residual battery level and required energy for reaching the predefined destination. Under energy trading among EVs, there is an energy pool managed by an auctioneer in the market that gathers surplus energy from sellers and supplies it to buyers using an auction mechanism. Each buyer individually decides the initial bidding price and the amount of energy to buy from sellers according to an expected cost savings compared with buying from macrogrids as well as its battery wear-out cost and charging/discharging efficiency. Similarly, each seller also individually decides its initial selling price and the amount of energy for sale subject to a tradeoff between the received revenue and discharging cost depending on its battery status. Notably, we provide a battery status model of EVs for designing well-defined utilities of both sellers and buyers. We design a naive auction process such that, in the auction process, the auctioneer controls the bidding increment to determine the best prices on a set of energies offered to multiple buyers through an iterative procedure. Finally, we show that distributing the energy based on a well-defined utility function converges to a unique optimal distribution for maximizing the payoff of all participating EVs.

For improving Quality of Service in wireless networks, multi-homing techniques have been considered as a promising solution. To date, most of the conventional research efforts have been focusing on user side interests (e.g. improving throughput, minimizing packet loss). However, a service provider could have different interests (e.g. profit, energy efficiency). Nevertheless, to the best of our knowledge, there is no research concern that takes into account the interests of user and service provider side at the same time. In this paper, we first model utility functions of both sides considering the aforementioned issues. Then, based on these utility functions we address the joint pricing and load distribution problem of multi-homing in heterogeneous wireless networks. Here, the problem is formulated into a Stackelberg game. Then, we propose schemes to obtain an optimal solution; so that, both sides (i.e. service providers and users) are satisfied. Finally, we provide rigorous analysis varying different parameters (e.g. cost, rate allocation, energy efficiency) which can potentially affect to the game. In addition to that, we show the proposed schemes are well converged to equilibrium point.

As the demand for high-quality multimedia service over wireless networks has increased wireless multicast communication has been researched for many years in efforts to achieve high throughput. Recently, cooperative multicast scheduling scheme was proposed with the goal of achieving high throughput with good fairness. However, although much attention has been directed towards energy efficient communication, energy efficient cooperative multicast design has not received much consideration. Part of the reason for this is that significant relay power consumption is required for user cooperation. In this paper, we model Quality of Service (QoS) of multicast in terms of outage probability when the cooperative multicast scheme is adopted. From the model, we obtain analytical results for QoS and power consumption under a varying number of users and cell coverage of the Base Station (BS). Finally, the analytical models can be used to find the optimal B S's transmission power by predicting system performance if system configurations are given. This can help system deployment and system optimization.

For multi-homed streaming service, it is important to enhance the throughput by efficient allocation of resources to the most appropriate interfaces of the User Terminal (UT). However, running multiple interfaces simultaneously can significantly contribute to rapid reduction of battery life. In this work, we propose a Power Minimized Rate Allocation Scheme (PMRAS) with optimal rate allocation at each interface with or without packet loss constraint. To develop the PMRAS, we formulate power consumption model based on the network initial interface state (e.g. active or sleep state). To solve convex optimization with multiple constraints, we adopt a Lagrangian algorithm based on dual decomposition. When compared with Rate Proportional additive-increase multiplicative-decrease (AIMD), the proposed algorithm gives significantly reduced total energy consumption with guarantee required Quality of Service (QoS) constraints.