Vehicular edge computing is rapidly amplifying the need to process computation-intensive tasks generated in vehicular environments. Conventional centralized processing frameworks struggle to meet these low-latency demands due to network latency and bandwidth constraints. Federated learning (FL) addresses these challenges by retaining raw data on vehicle nodes (VNs). However, sustainable FL training in vehicular edge computing faces persistent obstacles: the instability of VNs (high entry/exit rates), resource heterogeneity, and the lack of effective incentives. Notably, while existing incentive mechanisms are extensively studied, they inadequately address two critical sustainability barriers: extreme data heterogeneity and the volatile energy costs of mobile VNs. To overcome these challenges, we propose an auction-based sustainable federated learning (ASFL) framework. In this framework, the edge server acts as both the FL task publisher and the auctioneer, while VNs serve as bidders. Each bid encapsulates critical attributes including data quality, computing capacity, and projected energy cost. The core objective of ASFL is to maximize long-term social welfare. Formulating this objective reveals an inherently nonconvex optimization problem. Through rigorous analysis, we derive an equivalent convex formulation. The systematic bidder selection process inherent in ASFL simultaneously mitigates data heterogeneity and promotes rational energy utilization across FL. We theoretically prove that the framework achieves incentive compatibility and individual rationality. Experimental results on MNIST and CIFAR-10 datasets demonstrate the effectiveness of the method in mitigating the impact of non-i.i.d. data and reducing energy consumption.

Federated learning (FL) that preserves privacy has appeared as a prospective paradigm in mobile edge computing networks. However, due to the system and data heterogeneity of mobile clients (MCs), group-based FL with a sampling mechanism is crucial for minimizing model training costs. To address these challenges, we investigate and formulate the problem of group-based FL with a sampling mechanism for reducing model training cost (i.e., latency and energy consumption), and propose a group-based FL with a cost-efficient sampling mechanism (GFLCSM) framework to address it. More precisely, before training, each MC locally pre-trains a model, estimates its data distribution from the classifier's gradient norms, and uploads it to the central server (CS) instead of raw data to preserve privacy. Using this information, the CS transforms vanilla FL into a group-based FL. During training, GFLCSM replaces the random sampling mechanism with a cost-efficient one. Moreover, to enhance robustness against network dynamics, we extend GFLCSM with a backup resampling mechanism, termed GFLCSM-E. Experimental results indicate that GFLCSM surpasses the baseline frameworks, reducing latency by 24.63% and energy consumption by 11.47% on average across two datasets, while GFLCSM-E maintains high performance even under client dropout. The source code address ishttps://github.com/kt4ngw/GFLCSM.

Deploying federated learning (FL) in mobile edge computing (MEC) networks has become a prevalent approach to distributed learning. However, the inherent heterogeneity in computing, transmission and data on edge devices (EDs) brings challenges in improving training efficiency and speed. Existing approaches primarily focus on increasing batch sizes or employing adaptive batching to expedite convergence, but often overlook the generalization ability of the model. In this paper, we propose an energy-efficient adaptive batching approach for FL in MEC networks, aiming at minimizing the energy consumption by balancing training efficiency and speed. Initially, we exploit the relationship between batch size and loss improvement while determining the optimal learning rate corresponding to the batch size and understanding the correlation among loss improvement, learning rate, and gradient noise scale (GNS). Then we dynamically adjust the batch size based on the GNS and propose a low-complexity approach for measuring GNS. Finally, we fine-tune the batch size by assessing gradient similarity on each ED to ensure an optimal level of gradient noise during training, thereby enhancing the model's generalization. Experiment results demonstrate the effectiveness of our approach with an approximate 50% and 20% reduction in energy consumption and time consumption compared with existing approaches.

With the rapid deployment of large language model (LLM) technology across various fields, numerous LLMs have been deployed on edge servers (ESs) to provide low-latency generative services for edge devices. However, as factual knowledge evolves, the massive number of parameters in LLMs poses significant challenges for updating LLMs on ESs. Existing studies employ federated fine-tuning to update LLMs, but these unconstrained updating approaches can lead to overfitting and knowledge forgetting, while also resulting in substantial overhead. Knowledge editing (KE), as a promising technology, ensures the injection of new knowledge while preserving existing knowledge by editing specific parameters. In this paper, we propose multi-ES collaborative KE for the first time and design the CoKE and pCoKE frameworks in edge-cloud scenarios. These frameworks enhance editing efficiency by extracting identical expressions of the same knowledge across different LLM parameters for collaborative editing across multiple ESs. Additionally, we incorporate a personalized selection module in pCoKE to provide domain-specific generative services on ESs. To further reduce editing latency, we design a binary search-based resource allocation algorithm to balance editing latency across ESs. Extensive experiments demonstrate that CoKE and pCoKE reduce editing latency by 73% while maintaining high editing quality. Moreover, pCoKE achieves an approximately 6% improvement in editing quality.

In mobile edge computing (MEC) networks, federated learning (FL) has emerged as the leading distributed framework for training a shared machine learning model, primarily benefiting from its ability to exchange the information of edge devices (EDs) while safeguarding their privacy. However, in MEC networks, the heterogeneity of communication, computation, and data can result in challenges such as stragglers and data imbalances, thereby impeding the training process of FL. To address these challenges, we propose a Semi-Asynchronous Federated Learning (Semi-AFL) framework with cluster-based scheduling. In Semi-AFL, the EDs can perform local training at their own pace using different stale global models to tackle the straggler effect. Considering the asynchronousity of Semi-AFL and data heterogeneity, we propose a cluster-based scheduling strategy that includes device clustering and device selection. Specifically, it performs clustering based on the label distribution and obtains device-to-cluster information. We further select devices based on clustering information as well as model staleness and contribution, aiming to reduce variance and bias and accelerate model convergence. Experiment results demonstrate the effectiveness of the proposed method in reducing the latency of FL.

Computation offloading is widely used in vehicular edge computing (VEC) networks to satisfy the computational intensity and latency sensitivity requirements. However, many existing offloading algorithms do not comprehensively consider the dynamically changing characteristics of heterogeneous tasks within a roadside unit (RSU), resulting in tasks timeout and being dropped. In this paper, we design a competitive and cooperative computation offloading (C3O) model to reduce task execution latency. Specifically, when intensive heterogeneous tasks are generated, these vehicles jointly compete for the computing resource of a RSU, or they can also offload tasks to the task vehicle (TaV) with free computing resource. Meanwhile, We analyze the latency model of local execution and offloading to RSU or TaV execution and formulate a sequential task offloading decision problem, NP-hard. To address it, we propose a multi-agent reinforcement learning algorithm based on C3O (MARC3O) to intelligently determine the computation offloading policy for each vehicle according to the state of VEC networks. Simulation results demonstrate that the proposed algorithm can significantly reduce task execution latency and improve task completion rates compared with baseline schemes.

To break data silos and address the challenge of green communication, federated learning (FL) is widely used at network edges to train deep learning models in mobile edge computing (MEC) networks. However, many existing FL algorithms do not fully consider the dynamic environment, resulting in slower convergence of the model and larger training energy consumption. In this paper, we design a dynamic asynchronous federated learning (DAFL) model to improve the efficiency of FL in MEC networks. Specifically, we dynamically choose a certain number of mobile devices (MDs) by their arrival order to participate in the global aggregation at each epoch. Meanwhile, we analyze the energy consumption model of local update and upload update, and formulate the problem as a dynamic sequential decision problem to minimize the energy consumption, which is NP-hard. To address it, we propose an energy-efficient algorithm based on deep reinforcement learning named DDAFL, to intelligently determine the number of MDs participating in global aggregation according to the state of MEC networks at each epoch. Compared with baseline schemes, the proposed algorithm can significantly reduce energy consumption and accelerate model convergence.