Wide-area, time-sensitive Internet of Things service in sixth-generation non-terrestrial networks is constrained by latency, energy consumption and reliability. RLDS-NTN is presented as a mobile-edge-computing-assisted scheduling and control framework for LEO NTN-IoT systems. Belief–Desire–Intention extended agents are combined with deep reinforcement learning, federated model updates and uplink power-domain NOMA within a distributed decision loop. Resource, power and pairing actions are proposed by the DRL policy and are masked, validated or overridden by the BDIx layer under explicit congestion and feasibility rules. Two evaluation environments are considered. In the full-scale trace-driven campaign under noise-limited, clear-sky S-band assumptions, an SNR of 2.1 dB, a BER of 1.5×10−5, transmit power consumption of 1.26 W, Jain fairness of at least 0.96, packet loss below 0.4% and end-to-end latency of 38 to 46 ms are achieved. Component contributions are isolated in a MATLAB reference implementation under paired conditions. Delivered throughput is increased by 22% through NOMA and the latency tail is reduced. Residual loss from the unguarded learner is removed by the BDIx layer. Federated training matches or exceeds the local-only and centralized alternatives. A mean scheduler decision time of 5.2 ms with an 8.2 ms 99th percentile is measured on the documented x86 development platform, which satisfies the adopted 10 ms control period. These results are limited to the evaluated scenarios.
The rising prevalence of chronic disease and functional impairment in aging populations demands a shift from episodic care to continuous, remote monitoring. The Internet of Medical Things (IoMT) offers a promising solution for elderly and disabled individuals. This survey synthesizes IoMT-based systems across major chronic diseases —including diabetes, heart disease, chronic obstructive pulmonary disease (COPD), chronic kidney disease (CKD), arthritis, depression, dementia, and mobility impairments—analyzing them through a four-layer IoT architecture (perception, network, processing, application) and disease-specific comparison matrices that capture sensing, connectivity, analytics, validation, and deployment aspects. Findings reveal strong technical potential for continuous monitoring and early risk detection, yet significant heterogeneity in datasets, outcomes, validation protocols, and reporting practices limits comparability and translational confidence. Longitudinal real-world validation, usability evidence, and security-aware deployment remain scarce. To address these gaps, the paper proposes the Elderly-Centric IoMT Evaluation Framework (EC-IEF), which integrates four complementary dimensions: architectural-layer optimization, clinical and disease context, system design choices, and validation maturity with deployment readiness. By combining cross-disease synthesis with architecture-aware and deployment-aware analysis, this survey provides a unified basis for interpreting current IoMT solutions and identifying the technical, clinical, and operational priorities required for safe, scalable, and clinically meaningful adoption in elderly and disabled care.
This paper presents an Integrated Sensing and Communications (ISAC)-inspired hybrid fingerprinting framework for building-scale indoor localization on commodity smartphones. The framework adopts an opportunistic sensing paradigm, reusing existing 5G NR and Wi-Fi signals without dedicated waveform design. It fuses heterogeneous measurements, including 5G cellular metrics (RSRP/RSRQ/RSSI) and Wi-Fi RSSI from BSSIDs, into fixed-length descriptors. A supervised learning model then maps these descriptors to three-dimensional coordinates (x,y,z), where z encodes floor height.We make three contributions. First, we design a practical NR+Wi-Fi pipeline that lowers surveying overhead and targets deployments in campuses and hospitals. Second, we define a time-separated evaluation protocol: models are trained on an offline campaign and tested on measurements gathered later at the same anchors, probing robustness to ambient changes. Third, we compare DNN, CNN, and Transformer architectures under identical preprocessing and optimization schedules.Experiments in a three-story facility show that the Transformer achieves the best accuracy under the time-separated protocol. Across all floors, it attains a mean error of 4.6m and a median error of 4.2m, outperforming the CNN baseline (5.8m mean, 5.2m median) and the DNN baseline (7.1m mean, 6.5m median). This corresponds to mean-error reductions of approximately 21% vs. CNN and 35% vs. DNN, with similar gains for median error (about 19% and 35%, respectively). Cross-device tests with Google Pixel and Samsung handsets indicate device-robust operation without per-device calibration. A modality ablation further shows that the three 5G NR cellular metrics—just 1.6% of the 189-dimensional input—yield a 13% mean-error reduction over a Wi-Fi-only configuration, evidencing a disproportionately high per-feature contribution from the cellular modality.
The increasing popularity of containerized applications in edge and fog computing creates a strong need for high-quality guarantees of predictable real-time behavior. However, classical container orchestration is insufficient due to challenges such as dynamic workloads, heterogeneity, and performance interference caused by contention over shared hardware resources. Existing frameworks, including the original Hierarchical Resource Orchestration Framework (HROF), primarily rely on reactive control, use victim-based migration techniques that perform poorly against persistent interference sources, suffer from offline planning scalability limitations, and neglect vital non-CPU resources such as network bandwidth. This work presents the enhanced Hierarchical Resource Orchestration Framework (eHROF) to address these shortcomings. The eHROF employs proactive control strategies-through Model Predictive Control or Reinforcement Learning-to predict resource needs, an interference-aware migration strategy that identifies and manages disruptive containers using fine-grained shared resource monitoring, a scalable hybrid offline planner combining heuristics and targeted SMT/OMT solving, and intentional co-management of multiple resources (CPU and network). Systematic simulation experiments demonstrate that the eHROF achieves dramatically lower deadline miss ratios than baseline HROF and other approaches across multiple scenarios involving steady loads, high interference, dynamic workloads, and multi-resource constraints. By synergistically unifying proactiveness, interference awareness, scalability, and multi-resource support within a hierarchical control infrastructure, the eHROF provides a significantly more robust, predictable, and efficient platform for demanding real-time containerized workloads at the edge and fog.
Smart city waste management systems increasingly rely on distributed IoT sensor networks that collect real-time operational data, yet traditional access control mechanisms grant persistent credentials creating prolonged vulnerability windows and regulatory compliance challenges. Existing hierarchical access control schemes fail to address the dynamic, time-bounded nature of waste management operations, where collection drivers require temporary access to thousands of bins along fixed routes during strictly defined shifts. This paper presents WasteWhisperer, a Hierarchical Identity-Based Encryption framework specifically architected for ephemeral access control in eco-city cleanup operations. WasteWhisperer introduces novel wildcard pattern abstractions enabling cryptographically sound delegation that matches operational reality- drivers obtain access to "any bin on assigned route during shift'' rather than managing combinatorially explosive individual key sets. Through smart city-specific parameter tuning and wildcard pattern caching, WasteWhisperer achieves 50% performance improvement over state-of-the-art HIBE schemes while uniquely providing comprehensive capabilities including wildcard support, distributed storage, and fine-grained temporal control.
This paper proposes a novel Software-Defined Networking (SDN) architecture tailored for highly dynamic mobile networks, named SD-HDMN. To validate its feasibility and effectiveness, we introduce a lightweight network interaction protocol for highly dynamic scenarios, coupled with an efficient programmable forwarding scheme based on service feature recognition. Experimental results show that our lightweight network control protocol reduces overhead by 88% compared to existing solutions, while the corresponding link-state routing protocol achieves a 55.6% reduction. In a 112-node network, the system maintains an end-to-end latency of 35 ms with three controllers operating concurrently and reduces the network convergence time by 14.2%. Under a load of 4000 flow entries, the SD-HDMN scheme exhibits lower forwarding latency than a single-level flow table. Even in extreme scenarios, such as complete controller failure due to cyberattacks or physical destruction, the network performance remains comparable to that of a traditional distributed network, and data transmission is unaffected. The proposed SD-HDMN architecture demonstrates superior scalability, survivability, flexibility, and on-demand service capabilities over traditional distributed architectures.
The rapid proliferation of electric vehicles (EVs) highlights the necessity of accurate station-level charging demand forecasting for sustainable smart-city energy management. Reliable forecasts enable dynamic pricing, user guidance, and infrastructure planning, yet the heterogeneity of charging stations and the integration of multi-source inputs pose significant modeling challenges. To address these issues, we propose MSST-CDP, a multi-source spatio-temporal framework for fine-grained EV charging demand prediction. The spatial module employs a Graph Attention Network (GAT) to model pricing-driven inter-station dependencies, while the temporal module adopts a multi-phase encoder-decoder architecture integrating static attributes, historical observations, and known future covariates. Attention-based variable selection enhances model-intrinsic feature relevance and interpretability, and a demand-granularity-aware mechanism accounts for heterogeneous charger capacities. Experiments on a large-scale real-world dataset show that MSST-CDP outperforms multiple baselines across forecasting horizons and provides interpretable insights into spatio-temporal demand drivers for adaptive pricing and infrastructure planning.
The deep integration of communication technologies and the Internet of Vehicles (IoV) has accelerated the explosion of compute-intensive applications, such as autonomous driving perception, which are typically characterized as directed acyclic graph (DAG) based tasks with strict temporal dependencies. However, existing offloading strategies often neglect the unique chain blocking effect inherent in DAG tasks and struggle to cope with non-stationary bursty electromagnetic interference and resource competition at edge servers (ES) in vehicular environments. To address these challenges, this paper presents a collaborative optimization framework of task offloading, dynamic caching and resource allocation for DAG-dependent tasks, which is designed to resist complex channel interference. Firstly, a latest start time based priority quantization model (LST-PQM) is developed to explicitly define the execution order of dependent tasks and to avoid deadlocks when the coexistence of intra-task parallelism and inter-task resource competition are considered. Secondly, a dynamic programming based caching mechanism called CacheSelect is designed to achieve in-time pruning of DAG dependency chains through proactive reuse of intermediate computational results with high popularity, where the dependency bottlenecks between any two tasks are broken. Thirdly, the collaborative optimization process is transformed into a Markov decision process (MDP), and a hybrid deep reinforcement learning algorithm named PD-A2C-Offloading is proposed by integrating A2C, proximal policy optimization (PPO)-Clip mechanisms, and Dueling DQN architecture. Within this framework, the joint optimization of discrete offloading and continuous resource allocation is addressed, where the optimality of peak resource allocation (beta = 1) is strategically identified to mitigate the chain-blocking effect, and the convergence stability in stochastic interference environments is significantly improved through advantage function decoupling. Finally, experimental results in a complex simulation environment incorporating a Bernoulli-Rayleigh fading channel model demonstrate that the proposed scheme exhibits superior robustness under various vehicular densities and interference intensities compared to baseline algorithms such as PPO-Off, DQN-Off, DDQN-Off and GLS-Off, reducing the average task completion time by 12.7%-41.2%.
Centralised architecture models are often adopted in the development of smart cities to take advantage of the technologies provided by dominant companies such as Google, Meta, and Amazon. This approach often comes with significant drawbacks, such as (i) application developers becoming dependent on the policies and rules imposed by large companies, which are subject to change, and (ii) users having their data collected and used haphazardly, and often without authorisation, which can undermine the trust that users place in the solution. Infrastructure for smart cities should, therefore, incorporate decentralised technologies that satisfy or reinforce important quality attributes such as trustworthiness, privacy, and security. The main objective of this article is to survey and discuss how the services of a smart city can be implemented using decentralised architectures; we explain the advantages and disadvantages of decentralised architectures and emerging technologies. We review the specialised literature and compile a set of emerging technologies that are essential to conceive a decentralised architectural solution for smart cities. The current state-of-the-art in decentralisation shows that this model can be implemented with current technologies and help to satisfy the quality attributes mentioned above, and potentially reinforce the degree of trustworthiness of the solution. We also propose a four-dimension guideline that functions as a conceptual driver for structuring actions for researchers, system architects, policymakers, and city authorities in the design, evaluation, and long-term maintenance of decentralised smart cities, thereby enhancing their capacity to address escalating urban challenges. We conclude that the set of technologies discussed in this article has the potential to be reused and replicated across smart cities.
Accurate and continuous monitoring of human motion is essential for effective fall detection in Ambient Assisted Living (AAL) environments. However, achieving high-performance deep learning models depends significantly on optimizing hyperparameters. Bayesian Optimization has proven to be a powerful approach for efficiently exploring complex search spaces, enabling adaptive tuning to enhance both accuracy and computational efficiency. This paper presents a novel Bayesian Optimization framework for a hybrid model that integrates a Multi-Scale Graph Convolutional Network (MS-GCN) and a Bidirectional Temporal Convolutional Network (Bi-TCN) to enhance fall detection accuracy and efficiency. The Bi-TCN branch utilizes adaptive dilated separable convolutions with multiple dilation rates to capture short- and long-range dependencies, while the MS-GCN branch extracts hierarchical spatial features from human keypoint data. An attention mechanism further enhances the temporal feature fusion process. To optimize model performance, Multi-Objective Max-value Entropy Search with Trust-Region Bayesian Optimization (MES-TRBO) is integrated, which adaptively tunes hyperparameters by balancing classification accuracy and inference time. MES-TRBO dynamically expands or contracts the search region based on validation trends, ensuring efficient convergence for better performance. Moreover, Gaussian Noise regularization in both MS-GCN and Bi-TCN improves robustness against input variations and noise disturbances, ensuring stable feature learning. Extensive experiments on publicly available fall detection datasets, CAUCA Fall and UR Fall, demonstrate the effectiveness of the proposed method over conventional Bayesian optimization and ensemble learning approaches. Our model achieves 97.41% accuracy on the UR Fall dataset and 95.84% on the CAUCA Fall dataset, showcasing superior generalization and computational efficiency compared to baseline methods.
Accurate localization is essential for autonomous driving and Advanced Driver Assistance Systems (ADAS) to ensure safe and reliable vehicle navigation. In urban environments, multipath propagation is a major source of error in GNSS-based positioning, as signals reflect off buildings, structures, and other city elements, leading to degraded localization performance. This issue significantly impacts autonomous vehicles and ADAS applications, where precise positioning is critical for decision-making and safety. To address this challenge, we propose a method to simulate the multipath effect using a ray-tracing approach based on the 3D city model. The CARLA open-source simulator is used to recreate urban environments and vehicle trajectories, while an octree-based ray-tracing technique computes the possible signal reflections from GNSS satellites to vehicle positions. These signals are then used to estimate the amplitude and phase of both Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) components, allowing for the calculation of multipath-induced errors using standard Delay Lock Loop (DLL) mechanisms. Finally, raw pseudorange and carrier phase measurements are generated, incorporating the estimated multipath effects. The proposed model provides a more detailed simulation of multipath propagation and signal obstructions compared to existing methods. This is reflected in the increased positioning error observed in urban scenarios, particularly in areas with limited satellite visibility, such as under rail tracks. To promote open science and reproducibility, we also share the dataset obtained with this simulation pipeline, containing the generated measurements as well as the results presented in this paper. This simulation approach has broad applications, including the development of enhanced localization algorithms for ADAS and autonomous vehicles, multipath mitigation techniques, sensor fusion strategies, and Artificial Intelligence-driven positioning systems.
The integration of unmanned aerial vehicles (UAVs) with wireless power transfer (WPT) technology is promising for overcoming energy constraints in wireless rechargeable sensor networks (WRSNs). Earlier studies focused on UAV trajectory optimization, assuming full sensor charging at each visit, impractical in real-world scenarios. However, optimizing trajectory and hovering time considering constraints like limited UAV battery capacity and varying energy consumption rates among sensor nodes poses a significant challenge in developing efficient charging scheduling algorithms for WRSNs. This paper addresses the aforementioned challenges and introduces a periodic multi-node partial charging model for UAV-assisted WRSNs. It proposes an efficient charging schedule optimizing trajectory and hovering time to maximize harvested energy by sensor nodes. The charging scheduling problem is initially formulated as a mixed integer nonlinear programming problem, proven NP-hard. It is then reformulated as a monotone sub-modular maximization problem with partitioned matroid constraints, allowing a greedy solution approach to achieving a (1-1e) approximation. For large-scale networks, the differential evolution algorithm is employed. Simulation analysis demonstrates the proposed approach's performance improvement by maximizing harvested energy, reducing dead nodes, and optimizing UAV tour time.
Clustering high-dimensional sensor reports is a fundamental task in Mobile Crowd Sensing (MCS), yet enforcing differential privacy (DP) often degrades utility because DP noise on high-dimensional sufficient statistics grows with both feature sensitivity and dimension, and static per-iteration budgeting can waste a limited total privacy budget in iterative updates. This paper proposes SA-HDPCA, a sensitivity-aware DP clustering framework that targets more stable clustering under a fixed privacy budget through three tightly coupled components: (i) a Privacy-Utility Score (PUS) for feature filtering with optional PCA, which reduces the effective dimensionality while explicitly accounting for sensitivity cost; (ii) a one-shot DP proxy that enables robust initialization by performing meta-heuristic search only on a DP-protected low-dimensional proxy, reducing seed variance under privacy-induced uncertainty; and (iii) a feedback-driven privacy budget scheduler that reallocates per-iteration budgets based on stability signals computed from already released DP-noised centroids (post-processing), and thus requires no additional access to raw data and incurs no privacy loss beyond the accounted releases; its main effect is to reshape noise across iterations and improve stability rather than guarantee consistent utility gains. Experiments on two public high-dimensional sensing datasets (HAR and GAS) demonstrate that SA-HDPCA provides a stability-oriented, competitive privacy-utility trade-off under strict privacy budgets, remains competitive with classical baselines under matched accounting, stays above the PCA-DP reference at the core budgets under the same strict pipeline, and shows dataset-dependent trade-offs relative to recent DP clustering baselines used only for external positioning.
Mobile Ad Hoc Networks (MANETs) are highly dynamic and decentralized. This makes them vulnerable to security threats, such as Sybil, Blackhole, and misrouting attacks. This study proposes a hybrid Graph-based Anomaly Detection and Reinforcement Learning (GAD-RL) framework for real-time malicious node detection in MANETs. The approach builds temporal sequences of network graphs to capture both structural and behavioral features of nodes. Hybrid anomaly scores are computed by combining structural metrics, such as degree and centrality, with behavioral indicators, including packet delivery ratio and mobility. Detected anomalies are classified using a Deep Q-Network (DQN). Uncertain predictions are handled by a reinforcement learning (RL) fallback agent. Experiments are conducted on a combined synthetic MANET dataset and NSL-KDD data. The proposed method achieves 99.85% accuracy, 99.80% precision, 99.91% recall, and a 99.86% F1-score. The RL agent provides adaptive handling of uncertain cases. Comparative analysis shows that the proposed hybrid framework outperforms existing methods in accuracy and adaptability. The results confirm effective detection and attribution of attacks in dynamic MANET environments, supporting real-time deployment.
Large-scale battery-less backscatter sensor networks enable maintenance-free agricultural monitoring but operate under highly dynamic solar and ambient RF energy conditions. The absence of energy storage and uncontrolled node activation results in energy wastage, node outages, packet collisions, and reduced network throughput, necessitating accurate energy prediction and an adaptive scheduling strategy. This study aims to develop an energy-aware prediction and scheduling framework that enhances node uptime, communication reliability, and energy efficiency in battery-less backscatter sensor networks. A large-scale agricultural backscatter sensor network is simulated to generate realistic energy harvesting data. A Gradient-Guided Hippopotamus Optimization-based Energy-Aware Long Short-Term Memory (GGHO-E2-LSTM) model is developed to predict the future energy availability of each node. Based on these predictions, an energy prediction-driven greedy scheduling algorithm dynamically selects the top-K energy-sufficient nodes for sensing and backscatter transmission in each time slot. Simulation results show that the proposed GGHO-E2-LSTM prediction and energy-aware scheduling framework achieves high energy prediction accuracy with R2 of 0.99, MAE of 0.0028, improves node uptime to 89%, increases network throughput to 1540 packets/hour, and significantly reduces energy wastage and packet collisions compared to baseline schemes. These results confirm the effectiveness of the proposed prediction-aware scheduling framework in enhancing scalability, reliability, and energy efficiency of battery-less backscatter sensor networks.
The proliferation of Internet of Things (IoT) devices has increased the cyber-attack surface; however, traditional feature engineering for Intrusion Detection Systems (IDSs) frequently insufficiently represents class-specific data distributions and intricate feature interactions. This paper presents AICE-FE (Adaptive and Interactive Contextual Feature Engineering), an innovative framework that modifies features to improve class separability. AICE-FE comprises two different elements: Adaptive Anomaly-Weighted Scaling (AAWS), which adjusts numerical features according to their statistical differentiation between normal and attack categories, and Interactive Contextual State Encoding (ICSE), which generates a novel feature by evaluating composite states derived from various categorical attributes. AICE-FE, evaluated across three distinct IoT datasets, consistently enhances the performance of machine learning classifiers, increasing the F1-score by up to 3.7 percentage points (compared to the standard scaling baseline) for a Random Forest model on the WUSTL-EHMS-2020 dataset. By generating more discriminative and context-aware feature representations, AICE-FE offers a highly effective solution for securing complex IoT ecosystems.
