Wireless Sensor Networks (WSNs) have evolved rapidly and are widely deployed in various applications such as defence, healthcare, automation, and traffic monitoring in smart cities. Due to the rapid development of emerging technologies such as the Internet of Everything (IoE), large volumes of data are transmitted from numerous mobile sensor nodes through various routing mechanisms. In WSNs, efficient data transmission significantly influences energy consumption, node mobility, network topology, and scalability, making routing protocols a critical component in network design. However, many existing routing protocols suffer from high energy consumption, inefficient next-hop selection, and limited adaptability to dynamic network topologies caused by node mobility. These limitations reduce network lifetime and degrade data transmission performance in highly dynamic WSN environments. To address these challenges, this paper proposes a novel Cooperative Transmission and Clustering-based Dynamic Directional Routing Protocol (CTC-DDRP) for WSNs. This protocol operates in three different phases: neighbour transmission area discovery, clustering-based election and selection, and cooperative data transmission. The proposed routing protocol dynamically selects the next forwarding node based on residual energy and mobility criteria, while employing a duty-cycling mechanism to balance energy consumption among the deployed sensor nodes. This highly dynamic and cooperative routing scheme improves reliability and extends network lifetime, even in high-mobility scenarios. The performance of the proposed CTC-DDRP protocol is evaluated through Monte Carlo simulations and compared with several benchmark routing protocols. Simulation results clearly indicate that the proposed CTC-DDRP achieves a 16% higher packet delivery ratio, a 6.0 ms lower end-to-end delay, a 24.05 kbps higher throughput, 2.85 nJ energy savings per round, and 1896.5 nJ higher residual energy after 10,000 rounds, compared with other existing routing protocols under identical simulation conditions. These results indicate that the proposed CTC-DDRP protocol provides an efficient and reliable routing framework for next-generation WSN applications.
Passive indoor localization based on Wi-Fi probe requests (PRs) has attracted growing interest because it enables infrastructure-side spatial sensing without requiring explicit user participation. However, compared with conventional active Wi-Fi RSS fingerprinting, passive PR localization still lacks benchmark-ready open resources, shared end-to-end workflows, and controlled comparative evaluations under common conditions. This paper addresses that gap by presenting a unified end-to-end pipeline for passive Wi-Fi PR fingerprinting localization, together with an openly released localization-oriented dataset designed to support reproducible benchmarking. The proposed pipeline covers passive packet capture, centralized logging, preprocessing, fingerprint construction, model training, online localization, and performance assessment within a common experimental framework. Under this unified setting, representative fingerprinting paradigms are evaluated, spanning classical retrieval-based methods, ensemble learners, neural networks, and attention-based architectures. Rather than treating the study as a simple leaderboard comparison, the evaluation is structured to examine application-relevant trade-offs among localization accuracy, robustness, worst-case error, inference latency, and within-environment deployment sensitivity. The results show that, within the evaluated testbed, classical retrieval-based methods remain highly competitive for typical-case passive PR localization, while the most favorable attention-based configuration in this benchmark achieves the lowest observed maximum error among the evaluated methods without sacrificing runtime efficiency. The analysis further shows that monitor-node density and spatial receiver geometry materially influence localization performance within the same environment, with receiver geometry, in some cases, affecting performance as strongly as node count. Taken together, the study contributes a reproducible benchmark foundation for passive PR localization and clarifies the benchmark-level operational trade-offs that govern model selection under the present experimental setting.
In vehicular edge computing, service migration is challenging due to high vehicle mobility, rapidly changing wireless coverage and fluctuating edge resource availability. Although trajectory prediction has been introduced into migration decision making, existing studies commonly treat prediction and decision as loosely connected stages, where prediction outputs are used only as auxiliary features and the feasible action space induced by vehicle movement is not explicitly modeled. To address this issue, this paper proposes a service migration method that tightly couples map-aware motion prediction with deep reinforcement learning. Specifically, a Map-Aware Unscented Kalman Filter is developed to estimate short-term vehicle motion under road topology, road direction, speed limit and drivable region constraints, enabling proactive perception of coverage variation. Based on the predicted motion states, the candidate server set is dynamically constructed and an action masking mechanism is introduced to restrict migration decisions to feasible service nodes. An N-step Noisy Double Dueling Deep Q-Network is then designed to learn long-term migration policies in dynamic action spaces. Simulation results based on real urban vehicle trajectories show that the proposed method consistently outperforms baseline methods in average service latency, energy consumption, and energy delay product under different server scales, traffic intensities, and resource constrained settings. These results verify the effectiveness and robustness of the proposed method for vehicular service migration.
Driven by the rapid growth of 5G, gigabit broadband, mobile workforces, edge computing, and the Internet of Things (IoT), Wide Area Networks (WANs) have become increasingly cross-domain, heterogeneous, and dynamic. This evolution places greater demands on routing efficiency, load balancing, and Quality of Service (QoS). Traditional distributed routing lacks a global view, whereas centralized Software-Defined Networking (SDN) controllers may suffer from state latency and decision delays in large-scale multi-domain settings. To address these challenges, we propose SSRO, a cross-domain routing framework based on Multi-Agent Deep Reinforcement Learning (MADRL). By integrating dual-agent collaborative decision making with Graph Convolutional Networks (GCN), pheromone-inspired mechanisms, and Federated Learning (FL), SSRO supports dynamic and adaptive routing. Experiments on real-world topologies show that SSRO improves cumulative reward by up to 59.27%, reduces median maximum link utilization by up to 80.76%, and improves median throughput by up to 36.47% over the evaluated baselines.
Federated learning (FL) enables collaborative model training across resource-constrained Internet of Things (IoT) devices without directly sharing their private data. However, it is vulnerable to inference attacks, where attackers gain insights into private data from global model updates. Moreover, it suffers from the free-riding problem, where clients benefit from the global model without contributing any genuine local updates. Differential Privacy (DP) can be integrated into FL to tackle privacy leakage, resulting in differentially private FL (DPFL). Nevertheless, the interaction between DP and free-riding has not yet been systematically addressed. In this paper, we identify and characterize a fundamental three-way trade-off between DP noise, model utility, and detectability that constrains free-rider detection in DPFL. To address this, we propose Adaptive Noise-Floor-aware Detection (ANFD), which normalizes client update norms by an analytically derived DP noise floor, and applies robust median and Median Absolute Deviation (MAD)-based z-scoring with temporal smoothing to separate genuine clients from free-riders. Unlike existing baselines, ANFD preserves model utility by preferentially admitting genuine clients rather than excluding both groups, even if this requires tolerating some free-riders in the process. Simulation results on MNIST, Fashion-MNIST, CIFAR-10, and HAR datasets show that ANFD maintains up to a 4× higher test accuracy than the best-performing baseline (for instance, at DP noise σ=0.1 on MNIST), while keeping the genuine client exclusion rate below 0.4% across all datasets. These results validate our claims, showing that ANFD prevents catastrophic degradation of model utility as noise levels increase.
