In open underwater environments, ensuring accurate positions of sensors while protecting private information of localization systems presents a significant challenge. The physical channel differences between terrestrial and underwater networks render most existing privacy protection schemes designed for terrestrial networks inapplicable underwater. Moreover, limited research on underwater privacy protection has led to high implementation complexity and communication expenses. In this paper, to reduce the complexity of privacy protection, a secure mobile localization scheme using autonomous underwater vehicles (AUVs) as anchors is proposed for underwater sensor networks, based on adversarial neural cryptography utilizing acoustic channel features. Depending on whether eavesdroppers show interest in keys, two adversarial cryptography models are proposed to protect transmission of legitimate localization information and to actively counter eavesdroppers with learning capabilities in real time. Furthermore, to obtain effective keys and minimize unnecessary key transmission, random physical layer channel features are dynamically utilized as real-time keys for the cryptography system, and a synchronous channel probing protocol is designed for key generation. Simulation and experimental results demonstrate that, compared to other approaches, the proposed secure localization scheme effectively prevents the leakage of position information and maintains localization accuracy while operating at lower implementation complexity and communication expenses.

The criticality of efficient traffic forecasting in Intelligent Transportation System (ITS) has garnered significant academic attention. This study addresses the prevalent issue of distribution shift in real-world datasets, which often degrades performance, and explores the effectiveness of the channel-independence (CI), a technique recently proposed to mitigate this issue. While Spatio-Temporal Graph Neural Networks (STGNNs) are noted for their flexibility to represent road structures, their designs typically lack the capability to integrate CI without disrupting the spatial relationships, potentially limiting the performance. We present a novel approach that successfully integrates CI into spatial-temporal forecasting by incorporating distinct temporal, spatial, and predefined graph structure information within each channel. Moreover, STGNNs frequently emphasize intricate designs, which result in increased computational demands while offering only marginal improvements in accuracy. This paper presents ST-MLP, a streamlined spatio-temporal model constructed exclusively from cascaded Multi-Layer Perceptron (MLP) modules and linear layers. Experimental results indicate that ST-MLP outperforms numerous existing STGNNs in both accuracy and computational efficiency. Our findings advocate for further investigation into more streamlined and effective neural network architectures within spatial-temporal forecasting research.

As a prevalent field of study in machine learning, intelligent agents can perceive their surroundings and make informed decisions. In many research areas, such as autopilot systems, undersea explorations, and distributed robotics, researchers have traditionally employed unidirectional systems, which usually rely on perceptions from sensors to controllers, or end-to-end models, which generate actions directly from raw data. Nonetheless, in unidirectional systems, controller efficiency is intrinsically linked to sensor accuracy, which makes unidirectional systems lack a self-improving capability. Meanwhile, compared with functionally separated frameworks, end-to-end methods may have their own limitations in scalability, generality, interoperability, training costs, and so on. To fulfill this gap, we propose a new coordinated optimization framework for intelligent agents (COIA). We introduce an inverted optimization channel from controllers to sensors in traditional functionally separated frameworks through communication between devices, enabling closed-loop online evolutive learning (OEL). To the best of our knowledge, this article first presents a universal coordinated optimization framework among supervised learning and reinforcement learning (RL) models, without human labels or intervention. Our method allows heterogeneous agents to autonomously adapt to some special situations in open environments, which forms a basis of networked artificial general intelligence (AGI). We design an experimental paradigm of COIA with concrete cases, which shows a significantly large performance margin over unidirectional and end-to-end models. The performance margin grows with task complexity.

Next-generation networks are evolving into intelligent, programmable fabrics, changing the role of the network from a mere forwarding medium into an active participant in distributed task computation and decision-making processes. The deep integration of AI into the network – combined with programmable paradigms – is one of the main key enabling technologies in such a transformative leap. In this complex and highly heterogeneous context, static and fixed-located security measures become inadequate and ineffective. Recently, Agentic AI – that envisions the autonomous collaboration of AI agents – has been explored as a viable paradigm for self-organized autonomous networks. In this work, Agentic AI is extended beyond orchestration to propose an Agentic AI-based architecture for Security in 6G (AAIS-6G) in which security is treated as a native, dynamic, and context-aware capability: cooperating agents continuously perceive network conditions, reason over evidence and likely threats, act by synthesizing and deploying lightweight AI-based anomaly detection virtual network functions (VNFs), learn from outcomes, and collaborate across domains to co-evolve protection with network dynamics and the evolving threat landscape. A representative use case, involving a network of drones, demonstrates automatic, adaptive security reconfiguration at runtime, evaluating detection accuracy under evolving network conditions together with resource and orchestration overhead (CPU/energy and deployment latency).

This work addresses two essential components in the design of bus-based vehicular networks: a simulation model for bus mobility and a data forwarding algorithm. Mobility simulation models enable researchers to test and evaluate ideas in diverse, complex scenarios that would be impractical to assess in real-world settings due to time and cost constraints. However, developing simulation models that accurately capture the nuances of real-world mobility remains a challenging task that is widely studied in the literature. To this end, we use official data to introduce G2S, a simulation model framework for generating bus mobility on top of the well-known SUMO simulator. We provide three realistic simulation scenarios with varying traffic demands based on General Transit Feed Specification (GTFS) data from Greater Vancouver, Canada. From a data dissemination perspective, we propose BR4C (Bus Routing Protocol based on Contact, Community, and Centrality Characteristics), a historical-based data forwarding strategy designed to improve message delivery between bus lines. Our solution leverages knowledge extracted from past encounters between buses, incorporating metrics such as community structure, centrality, and contact characteristics to enhance message forwarding. BR4C significantly reduces delivery latency while achieving a higher delivery ratio compared to stateof- the-art approaches.

The mobile edge computing (MEC) is a key technology for enabling energy-efficient and low-latency processing of computation-intensive applications through task offloading. However, current frameworks typically model-dependent tasks using static directed acyclic graphs (DAGs), which are poorly suited to dynamic edge environments characterized by fluctuating resources and diverse quality-of-service (QoS) demands. These fixed DAG structures often fail to adapt to runtime changes in bandwidth or node workload, leading to frequent task failures or inefficient resource usage. To overcome these limitations, we propose dynamic application configuration, a paradigm that allows applications to switch between multiple predefined DAG variants at runtime. This enables the system to dynamically balance accuracy and resource efficiency at critical decision points by adapting to current network and computing conditions. Based on this concept, we design the dynamic application configuration and capacity-aware task offloading system (DACCS), which employs a two-step offloading strategy: 1) a dynamic graph selection (DGS) algorithm that adaptively adjusts application configurations during key subtask execution based on real-time resource states and 2) a dependency-aware offloading algorithm (DAP) that optimizes offloading assignments by jointly optimizing the application completion rate, processing delay, and effective utilization rate. Experiments and real-system validations demonstrate that DGS improves the service performance across multiple strategies. Furthermore, the proposed DGS-DAP strategy outperforms other benchmark approaches under dynamic conditions.

Computer vision embedded in Internet of Things (IoT) systems will enable a new era of smart applications where video inference provides contextual awareness for the system. The limited resource capabilities of IoT devices and edge computing servers, often used to support computation-intensive IoT tasks, might not be enough to process video content produced by IoT devices in a computer vision-based smart system. In contrast to state-of-the-art where IoT video inference is performed locally at IoT devices or at edge and cloud servers, we propose a novel collaborative IoT paradigm where IoT devices share their idle resources for the processing of video frames in video analytics systems. We proposed a novel stochastic framework for modeling scenarios of collaborative IoT and edge/cloud continuum for video analytics systems. The proposed mathematical framework considers the unique characteristics of video analytics systems, IoT devices, and edge and cloud servers used to process video flows from IoT cameras in a collaborative manner. The obtained results show that the collaborative processing at neighboring IoT devices, i.e., IoT helpers, contributes to reduce the overall latency for video inference. However, high offloading costs might becoming a limiting factor which would request the design of more efficient offloading strategies.

The growing adoption of multimedia sensing has led to a rapid increase in data volume, urging the development of more efficient communication paradigms to alleviate transmission pressure. Considering the coarse nature of images and videos, only limited portions of them are required to capture key information and perform downstream tasks. Especially under strict bandwidth and latency constraints, compression based on semantic understanding represents a promising solution. However, the potential of leveraging vision–language models (VLMs) to extract image semantics for reducing communication burden remains underexplored. In this paper, we propose mask-aware semantic communication (MASC), a VLM-integrated framework in which a lightweight VLM deployed at the edge extracts a compact semantic representation (e.g., a set of labels or scene elements) under an explicit token/bit budget and transmits compressed data guided by a semantic mask. Experimental results demonstrate that MASC outperforms traditional compression methods and comparable semantic approaches under low signal-to-noise ratio (SNR) conditions, achieving robust reconstruction and high perceptual quality over additive white Gaussian noise (AWGN) channels.

This paper presents a new routing algorithm designed to tackle congestion in urban road networks, using a framework called Maximum Nash Welfare (MNW) that focuses on indivisible goods. Unlike traditional methods that simply find the shortest routes for individual drivers, our approach uses the MNW fairness and efficiency principles to distribute vehicles more evenly across different paths, especially when congestion is identified. We conducted simulations using the SUMO traffic model, and the results show that our method significantly reduces average travel times, reduces queue lengths, and minimizes idle times compared to conventional routing strategies. In this way, we are able to strike a balance between fairness, individual user satisfaction, and the overall efficiency of the road network.

This paper addresses the Traffic Signal Control (TSC) problem in urban road networks, where multiple distributed intersections must coordinate their signal scheduling under incomplete observation conditions. As is well known, realworld distributed systems often suffer from partial node failures, sparse sensor deployments, and similar issues, leading to significant information asymmetry across the control network. In TSC systems, these classical problems in distributed systems typically manifest as missing traffic data, which results in inefficient traffic signal control, exacerbating fuel waste through excessive stop-and-go movements and idling. This issue of missing data remains largely unaddressed by most Deep Reinforcement Learning (DRL) methods, which generally assume full observability. Moreover, existing approaches that tackle missing data in TSC often overlook the optimization of distributed fuel efficiency. To address the challenges mentioned before and the corresponding fuel consumption issue, this paper proposes a fuel-economic reinforcement learning model for distributed traffic signal control with spatiotemporal decomposition, called FEDLight. Our approach introduces a principled spatiotemporal decomposition mechanism that enables each intersection to locally compute fuel efficiency indicators while preserving global coordination. By integrating an adaptive data imputation module, FEDLight can effectively handle heterogeneous missing data patterns without requiring a centralized data processor. Extensive experiments on large-scale urban traffic networks demonstrate that FEDLight outperforms centralized approaches and other advanced models in reducing fuel consumption while exhibiting strong robustness to varying levels of missing data. This makes it well-suited for deployment in resource-constrained distributed environments.

Cloud/edge-native computing is a novel paradigm designed to reduce the deployment, management, and orchestration complexities of monolithic systems. This paradigm breaks down complex systems into a collection of microservices to fulfill applications' requirements and it is becoming popular in 6 G network-based systems, such as in envisioned use cases of mobile immersive reality. When following the cloud/edgenative computing paradigm, the pipeline of immersive reality systems is implemented through a collection of microservices deployed over the cloud/edge continuum. One challenge is to find microservice compositions capable of fulfilling the user's application requirement. This is done by determining the possible collection of microservices composition capable of handling user's demands. While several studies investigated the deployment of microservices and the automated generation of service compositions, less focus has been devoted to the convergence time of designed algorithms, which is critical given the time-sensitive nature of 6 G envisioned use cases. This paper designs a novel algorithm for service composition in edge-native computing. The proposed algorithm implements a memoization component to reduce search space when determining compatibility service graphs. Preliminary results show that the proposed algorithm outperforms traditional depth first search (DFS) by 87 % in terms of processing time while finding the same compatibility service graphs.

Edge-cloud collaborative computing (ECCC) has emerged as a pivotal paradigm for addressing the computational demands of modern intelligent applications, integrating cloud resources with edge devices to enable efficient, low-latency processing across distributed communication networks. Recent advancements in AI, particularly deep learning and large language models (LLMs), have dramatically enhanced the capabilities of these networked systems, yet introduce significant challenges in model deployment, network resource management, and cross-layer optimization. In this survey, we comprehensively examine the intersection of distributed intelligence and model optimization within edge-cloud environments, providing a structured tutorial on fundamental architectures, communication protocols, and network-aware computing frameworks. Additionally, we systematically analyze model optimization approaches, including compression, adaptation, and neural architecture search, alongside AI-driven resource management strategies that balance performance, energy efficiency, and communication overhead across heterogeneous networks. We further explore critical aspects of privacy protection and security enhancement within ECCC systems and examine practical deployments through diverse networked applications, spanning autonomous driving, healthcare, and industrial automation. Performance analysis and benchmarking techniques are also thoroughly explored to establish evaluation standards for these complex distributed systems. Furthermore, the review identifies critical research directions including LLMs deployment, 6G integration, neuromorphic computing, and quantum computing, offering a roadmap for addressing persistent challenges in heterogeneity management, real-time processing, and scalability. By bridging theoretical advancements in communications with practical deployments, this survey offers researchers and practitioners a holistic perspective on leveraging AI to optimize distributed computing environments over next-generation communication networks, fostering innovation in intelligent networked systems.

Underwater Wireless Sensor Networks (UWSNs) are essential for oceanographic monitoring, marine exploration, and disaster prevention. However, their performance is severely affected by high propagation delays, dynamic topologies, energy constraints, and void regions that disrupt data transmission. To overcome these challenges, this paper proposes TAVAR, a Topology-Aware and Void-Avoidance Routing protocol based on Genetic Algorithms (GA). TAVAR introduces a novel topology-aware initialization method that generates a high-quality initial population, significantly improving convergence speed and routing efficiency. It further employs an adaptive crossover strategy and dynamic scaling function to enhance the global search capability of GA in complex underwater environments. A multi-factor fitness function is designed by incorporating local topology, node reliability, and link stability to support robust and adaptive routing decisions. Moreover, to address the communication void problem, an intelligent depth adjustment mechanism is integrated to dynamically reposition void nodes with minimal energy overhead, thereby restoring connectivity and reducing packet loss. Simulation results demonstrate that TAVAR achieves superior performance compared to existing protocols in terms of packet delivery ratio, end-to-end delay, and energy efficiency, making it a promising solution for robust and energy-aware routing in UWSNs.

Computer vision is an enabling technology for the next generation of intelligent systems. The timely inference of video frames produced by mobile devices will provide contextual awareness through visual information to determine the course of action of mobile autonomous robots or enhance the physical world perception of mobile users, for instance. One of the fundamental challenges in mobile video analytics systems is the efficient management and allocation of computing resources to video inference requests. Edge computing emerges as a promising solution for providing compute resources for video analytics. However, it has limited resources compared to cloud servers, which requires efficient management and allocation. In this paper, we propose a stochastic framework to model a mobile video analytics system where networked edge servers are used to support video frame inference. We formulate the problem of deciding whether incoming video frame inference requests should be served or blocked at edge servers, as well as whether inference tasks currently being served need to be migrated to another edge server, while minimizing the request block rate. Furthermore, we propose a novel deep reinforcement learning (DRL)-based framework for allocating appropriate edge servers to tasks originating from different regions with varying computational and latency requirements, and for migrating tasks during the service period as needed. Numerical results show that our proposed framework effectively balances workload at edge servers, reduces task block rate, and can timely identify changes in regional load arrivals based on historical information, using the migration mechanism to prepare accordingly.

Video inference will be a building block in many of the envisioned mobile smart systems in 6G networks. Mobile devices will produce video frames that will be analyzed in real-time to enhance users’ perception of the physical world or determine actions to be taken by mobile autonomous robots and smart systems. However, video inference on mobile devices is challenging due to the devices’ limited resources, whereas inference on cloud servers is daunting due to constrained uplink bandwidth. Therefore, communication and processing resources must be efficiently managed and allocated jointly to meet stringent video inference requirements in smart mobile systems. In this paper, we propose a cross-layer deep reinforcement learning (DRL) framework that jointly optimizes the offloading rate at mobile devices, the allocation of uplink physical resource blocks (PRBs), and the orchestration of multi-DNN video frame inference on edge servers. We devise a stochastic framework that models the dynamics of distributed video frame analytics across devices and edge servers, as well as the dynamic optimization of uplink wireless communication links and GPU sharing for multi-DNN inference at edge servers. Furthermore, we formulate the joint optimization of wireless uplink communication and edge compute resources as a Markov decision problem and propose a recurrent deep reinforcement learning framework to optimize communication and computation resource allocation under varying uplink channel quality, demand for computation resource, and video frame inference latency and accuracy requirements. Numerical results show that the proposed framework improves inference accuracy and robustness while maintaining real-time latency and outperforming baseline solutions.

In IoT resource-constrained systems, task offloading to edge and/or cloud servers is the traditional approach to handle computationally intensive tasks. However, this classical approach might not be suitable for coping with artificial intelligence (AI) tasks of the next generation of intelligent and time-critical IoT systems. Recently, deep neural network (DNN) splits have been proposed to handle IoT AI tasks on different devices. Nevertheless, this approach might incur high processing and communication costs when the proper DNN split is not achieved. Therefore, this paper proposes a novel mathematical framework to model the performances and trade-offs of DNN splitting and intelligent offloading techniques in edge-aided, computer vision-based IoT intelligent systems. The proposed framework captures the unique characteristics of task offloading and DNN split process for IoT systems, including task characteristics and QoS requirements, layer-wise DNN inference, data transmission, and task computation. Besides, we model the end-to-end latency, edge computing resources’ utilization, and video frame drop rate when DNN splitting and intelligent offloading techniques are employed to alleviate the video inference burden in resource-constrained IoT devices. The obtained results show that the proper selection of the DNN splitting point and offloading ratio is critical to optimizing the performance of IoT video inference. Splitting at pooling layers can reduce the offloading data size without incurring excessive local latency, while intelligent offloading outperforms DNN splitting at low workloads. The obtained findings provide practical guidelines for configuring offloading strategies in real-world edge-assisted IoT video analytics deployments.

Fountain codes with online adaptation (OFCs) are promising for underwater acoustic sensor networks (UASNs), since they exploit limited feedback to reduce transmission over head. Yet, the performance of conventional OFCs is severely degraded in UASNs due to high error rates, long propagation delays, and sparse feedback, resulting in poor recovery efficiency and high energy cost. To address these challenges, we propose an energy-aware dual-perception OFC framework driven by deep reinforcement learning (DRL-OFC). In this design, a DRL agent jointly perceives channel dynamics and feedback sparsity to learn an optimal degree distribution that suppresses redundant coding and enhances intermediate decoding. In addition, a feedback aware transmission strategy is developed to cope with the long delay characteristics of UASNs, further reducing unnecessary retransmissions. Simulation results show that DRL-OFC achieves significant gains over existing OFC schemes in terms of decoding efficiency, transmission overhead, recovery performance, and energy consumption, confirming its suitability for resource constrained underwater communications.

In harsh underwater environments, the localization of network nodes faces severe challenges due to open deployment environments. Most existing underwater localization methods suffer from privacy leaks. However, privacy protection schemes applied in terrestrial networks are not viable for underwater acoustic networks due to stratification effects and multipath complexities. In this paper, we introduce a secure localization scheme for underwater wireless sensor networks (UWSNs) utilizing cooperative beamforming among mobile underwater anchor nodes. With this scheme, the underwater sensor communicates and ranges with mobile anchor nodes to perform self-localization via time difference of arrival (TDOA) algorithm. However, the presence of eavesdroppers poses a threat by intercepting information emitted by the anchors. To avoid localization information leakage, then we model the secure localization requirement as a multi-anchors multi-objective dual joint optimization problem to enhance both security and energy performance. The deep reinforcement learning (DRL)-based multi-agent deep deterministic policy gradient (MADDPG) algorithm is applied to solve the optimization problem. Both simulation and field experimental results robustly validate the efficiency and accuracy of the proposed secure localization scheme.

The growing demands for low-latency and secure communication in vehicular networks necessitate reliable mobility management protocols. In this paper, we propose a proactive, hierarchical mobility management protocol designed to enhance the security of vehicular networks during handover processes. Our protocol emphasizes early registration and pre-authentication to mitigate risks associated with rapid vehicle mobility, such as impersonation, replay attacks, and data breaches. We evaluate the protocol’s effectiveness against common security threats in mobility management using real-world mobility traces.

Accurate traffic flow forecasting is essential for various traffic applications, such as real-time traffic signal control, demand prediction, and route guidance. However, the increasing complexity and non-linearity of big data in the traffic domain pose a challenge for accurate forecasting, necessitating powerful models. This paper proposes a Spatio-temporal model for traffic flow prediction based on Graph Convolutional Neural Network (GCN) and Convolutional Neural Networks (CNN). The hierarchical architecture of Spatio-temporal modeling is utilized to consider multi-scale Spatio-temporal dependencies. We evaluate the proposed model using three real-world datasets, including METR-LA, PeMS04(S), and PeMS04(L). Our experiments demonstrate that the model captures comprehensive spatiotemporal correlations with multi-scale semantics, outperforming features extracted from single domains and non-multi scales. Furthermore, the proposed model is powerful for long-term prediction. We also conduct ablation and architecture studies to highlight the importance of model architecture for Spatiotemporal feature extraction. Our proposed Spatio-temporal model based on GCN and CNN offers a promising approach to traffic flow forecasting in complex traffic scenarios.