The Cloud Continuum Framework (CCF) extends computing capabilities across near-edge, far-edge, and extremeedge nodes beyond the traditional edge to meet the diverse performance demands of emerging 6G applications. While Deep Reinforcement Learning (DRL) has demonstrated potential in automating Virtual Network Function (VNF) migration by learning optimal policies, centralized DRL-based orchestration faces challenges related to scalability and limited visibility in distributed, heterogeneous network environments. To address these limitations, we introduce MARC-6G (Multi-Agent Reinforcement Learning for Distributed Context-Aware Service Function Chain (SFC) Deployment and Migration in 6G Networks), a novel framework that leverages decentralized agents for distributed, dynamic, and service-aware SFC placement and migration. MARC-6G allows agents to monitor different portions of the network, collaboratively optimize network control policies via experience sharing, and make local decisions that collectively enhance global orchestration under time-varying traffic conditions. We show through simulations that MARC-6G improves SFC deployment efficiency, reduces migration costs by $\mathbf{3 4 \%}$, and lowers energy consumption by $\mathbf{1 2. 5 \%}$ compared to the state-of-the-art centralized DRL baseline.

Advanced beamforming techniques enable stable vehicular communication and address mmWave limitations by accurately directing the signal. However, traditional beamforming techniques struggle in high-speed vehicles due to time-intensive codebook processing and image-based feedback adjustments. Multi-modal beamforming using real-time data like GPS, cameras, and LiDAR to train the Deep Learning (DL) models can provide adaptive beam steering, improving reliability in dynamic conditions. Despite this, centralized systems involving large raw data transmission are vulnerable to saturation and malicious interference, and they neglect privacy concerns, necessitating a new framework. This paper proposes a novel federated attentionbased fusion learning framework named FedAttention for multimodal beamforming in the Internet-of-Vehicle (IoV). FedAttention further improves the model generalization ability by utilizing the CNN-Transformer architecture and making full use of the Multi-access Edge Computing (MEC) servers for the potential federated split learning to enhance efficiency. Based on the realworld datasets, FedAttention achieves 98.16 % in Top-5 accuracy and 82.09 % in Top-1 accuracy, a 26.86 % improvement compared to the current FLASH framework with less wall clock time, showing its training efficiency and robustness.

In Vehicle-to-Everything (V2X) communication, advanced beamforming techniques address signal attenuation caused by mmWave, which provides high bandwidth and low latency. Multi-modal beamforming using Federated Learning (FL) can leverage resources like GPS, Lidar, and image data, significantly accelerating beam searching while enhancing data privacy. The heterogeneity of vehicles, however, affects the availability of computing resources for training machine learning models. Moreover, the multi-modal fusion network may contain billions of parameters, leading to extended training time for FL. To address these challenges, this paper proposes a novel Deep Reinforced Federated Split Learning framework (DRFSL) tailored for multi-modal beamforming with different sub-model architectures. DRFSL efficiently utilizes MEC computing and adapts the collaborative and distributed training to dynamic network conditions and system heterogeneity by incorporating deep reinforcement learning and split learning with FL. Experimental evaluation using real-world datasets demonstrates that DRFSL minimizes average training time by 49.45% and inference time by 24.43% and can achieve higher accuracy within the same timeframe compared to the existing FLASH framework.

Federated Learning (FL) is widely applied in privacy-sensitive domains, such as healthcare, finance, and education, due to its privacy-preserving properties. However, implementing FL in dynamic wireless networks poses substantial communication challenges. Central to these challenges is the need for efficient communication strategies that can adapt to fluctuating network conditions and the growing number of participating devices, which can lead to unacceptable communication delays. In this article, we propose Stochastic Client Selection for Tree All-Reduce Federated Learning (CSTAR-FL), a novel approach that combines a probabilistic User Device (UD) selection strategy with a tree-based communication architecture to enhance communication efficiency in FL within densely populated wireless networks. By optimizing UD selection for effective model aggregation and employing an efficient data transmission structure,CSTAR-FLsignificantly reduces communication time and improves FL efficiency. Additionally, our approach ensures high global model accuracy under scenarios where data distribution is heterogeneous from User Device (UD)s. Extensive simulations in dynamic wireless network scenarios demonstrate thatCSTAR-FLoutperforms existing state-of-the-art methods, reducing model convergence time by up to 40% without losing the global model accuracy. This makesCSTAR-FLa robust solution for efficient and scalable FL deployments in high-density environments.

The key impediments to deploying deep neural networks (DNN) in IoT edge environments lie in the gap between the expensive DNN computation and the limited computing capability of IoT devices. Current state-of-the-art machine learning models have significant demands on memory, computation, and energy and raise challenges for integrating them with the decentralized operation of heterogeneous and resource-constrained IoT devices. Recent studies have proposed the cooperative execution of DNN models in IoT devices to enhance the reliability, privacy, and efficiency of intelligent IoT systems but disregarded flexible finegrained model partitioning schemes for optimal distribution of DNN execution tasks in dynamic IoT networks. In this paper, we propose DISNET, a distributed micro-split deep learning scheme for heterogeneous dynamic IoT. DISNET accelerates inference time and minimizes energy consumption by combining vertical (layer-based) and horizontal DNN partitioning to enable flexible, distributed, and parallel execution of neural network models on heterogeneous IoT devices. DISNET considers the IoT devices’ computing and communication resources and the network conditions for resource-aware cooperative DNN Inference. Experimental evaluation in dynamic IoT networks shows that DISNET reduces the DNN inference latency and energy consumption by up to 5.2× and 6×, respectively, compared to two state-of-the-art schemes without loss of accuracy.

The deployment of Internet of Things (IoT) temperature sensors in urban areas is essential for the monitoring and understanding of the thermal environment. However, accurate temperature measurements can be compromised by factors such as direct sunlight, leading to overheating and inaccurate readings. We propose a Machine Learning-based approach that addresses this challenge by dynamically ventilating the sensor environment using small fans, enabling accurate and energy-efficient temperature measurements. This paper focuses on two interconnected problems: predicting steady-state temperature using a limited window of initial temperature measurements and investigating the impact of ventilation time. We employ various DNNs suitable for low-power IoT sensor devices to predict temperature using multivariate time series from different sensors and compare their accuracy. Furthermore, we highlight the tradeoff between prediction accuracy, which is correlated to the length of the observed input sequence, and energy consumption dependent on ventilation time. By adopting advanced prediction techniques, we can develop efficient IoT systems for accurate and energy-efficient environment monitoring in smart cities.

Large Machine Learning (ML) models require considerable computing resources and raise challenges for integrating them with the decentralized operation of heterogeneous and resource-constrained Internet of Things (IoT) devices. Running ML tasks on the cloud can introduce network delay, throughput, and privacy concerns, whereas running ML tasks on IoT devices is penalized by their constrained resources. For this reason, recent research proposed cooperative execution of ML tasks over IoT networks but disregarded resource variability and the IoT devices’ energy constraints simultaneously. In this paper, we propose Early Exit of Computation (EEoC), an adaptive, energy-efficient, low-latency inference scheme over IoT networks. EEoC adaptively distributes the inference computation load between the IoT device and the edge server, based on estimated communication and computation resources, to jointly minimize prediction latency and energy consumption. We evaluate our solution’s latency and energy profile on a real testbed running two widely used neural networks. Results show that EEoC can reduce latency and energy consumption up to 24.6% and 46.5%, respectively, compared to other state-of-the-art solutions without sacrificing accuracy.

Distributed training of Machine Learning models in edge Internet of Things (IoT) environments is challenging because of three main points. First, resource-constrained devices have large training times and limited energy budget. Second, resource heterogeneity of IoT devices slows down the training of the global model due to the presence of slower devices (stragglers). Finally, varying operational conditions, such as network bandwidth, and computing resources, significantly affect training time and energy consumption. Recent studies have proposed Split Learning (SL) for distributed model training with limited resources but its efficient implementation on the resource-constrained and decentralized heterogeneous IoT devices remains minimally explored. We propose Adaptive REsource-aware Split-learning (ARES), a scheme for efficient model training in IoT systems. ARES accelerates training in resource-constrained devices and minimizes the effect of stragglers on the training through device-targeted split points while accounting for time-varying network throughput and computing resources. ARES takes into account application constraints to mitigate training optimization tradeoffs in terms of energy consumption and training time. We evaluate ARES prototype on a real testbed comprising heterogeneous IoT devices running a widely-adopted deep neural network and dataset. Results show that ARES accelerates model training on IoT devices by up to 48% and minimizes the energy consumption by up to 61.4% compared to Federated Learning (FL) and classic SL, without sacrificing model convergence and accuracy.