Rapid urbanisation, agricultural intensification, and climate variability have accelerated the deterioration of groundwater and surface-water quality worldwide. Traditional assessment methods—largely dependent on laboratory analysis and deterministic modelling—are often time-consuming, data-limited, and incapable of capturing the nonlinear interactions governing hydro-chemical processes. In response, Artificial Intelligence (AI) and Machine Learning (ML) techniques have emerged as transformative tools for water-quality prediction, classification, and decision support.This survey synthesises findings from over forty peer-reviewed studies (2019–2025) covering statistical, geostatistical, supervised, unsupervised, and deep-learning approaches applied to groundwater and surface-water datasets. We systematically compare our scope and contributions with four major existing surveys, demonstrating that this review uniquely integrates (i) a five-family taxonomy spanning statistical methods through hybrid-optimisation frameworks, (ii) an explicit communication-layer perspective linking IoT sensor networks, edge computing, and green ICT to water-quality intelligence, and (iii) applied explainable-AI (XAI) and physics-informed learning analysis absent from prior work.The review evaluates model architectures including Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN), Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and hybrid neuro-fuzzy and genetic-optimisation frameworks. Reported accuracies frequently exceed 95%, with ensemble and deep models outperforming classical statistical regressions in predicting the Water Quality Index (WQI) and contamination risk. A concrete prediction-task formulation distinguishes regression from classification and anomaly detection, specifying typical input features, temporal horizons, and evaluation protocols. Beyond accuracy, emerging trends emphasise explainability (SHAP, LIME), feature selection (entropy-correlation, PCA), and spatial integration with GIS and remote sensing. Despite remarkable advances, persistent challenges include data imbalance, regional transferability, and limited interpretability of complex models. The review concludes by outlining future research needs in interpretable hybrid frameworks, physics-informed learning, and real-time AI–IoT water-quality management.
Unmanned aerial vehicles (UAVs) have emerged as an effective platform for disaster monitoring due to their ability to capture high-resolution imagery in complex and hard-to-reach environments. However, accurate disaster recognition remains challenging due to distributed data collection and heterogeneous data distributions across UAVs. In this work, we present a Federated Vision-Language Prompting (FedVLP) for UAV-based disaster recognition that enables multiple UAV clients to collaboratively train a shared model without centralized data aggregation. To improve robustness under imbalanced and non-IID data conditions, the proposed approach incorporates class-imbalance aware training and a Dirichlet-based data partitioning strategy. Furthermore, a vision-language prompting mechanism is introduced to enhance feature representations by aligning visual features with class-level semantic descriptions, improving discrimination between visually similar disaster categories. A lightweight backbone network is employed to ensure computational efficiency for resource-constrained UAV platforms. Experimental results on the AIDERv2 and DisasterEye datasets demonstrate that the proposed framework achieves superior performance compared with strong baselines while maintaining a favorable efficiency–accuracy trade-off. These findings highlight the effectiveness of integrating federated optimization with semantic guidance for scalable UAV-based disaster recognition.
Fog computing is gaining prominence as a paradigm to support latency-sensitive applications by bringing computation closer to data sources. However, its decentralized and distributed nature makes it susceptible to trust-based attacks, especially ballot stuffing, where colluding nodes inflate the trust of a malicious peer. Traditional trust models based on weighted averages lack adaptability to dynamic, malicious behaviors and fail to distinguish between genuine and manipulated recommendations. This paper proposes FLAME, a Fuzzy Logic-Based Anomaly Model that integrates exponential penalization to detect inconsistencies between recommendation and contextual trust. By computing a dynamic Fuzzy Trust Score (FTS), FLAME detects malicious nodes even when their recommendation trust is artificially high. The model assigns linguistic trust labels (Low, Medium, High) using fuzzy membership functions and penalizes mismatches using a tunable λ parameter. Simulations on a Byzantine fault-tolerant network demonstrate the effectiveness of FLAME in accurately identifying ballot stuffing behavior, outperforming static trust models.
The concept of Internet of Medical Things (IoMT) has improved healthcare delivery by enabling continuous patient monitoring. The inherent nature of IoMT (e.g., heterogeneous and privacy-sensitive) introduces significant security vulnerabilities. Traditional intrusion detection approaches face limitations in scalability and data privacy. To address these challenges, a novel three-phase heterogeneous Federated Intrusion Detection System (FIDS) is presented that integrates federated representation learning and gradient-boosted decision trees in a privacy-aware ensemble framework. The proposed framework is evaluated on the CIC-IoMT 2024 WiFi/MQTT dataset (8.775 million samples, 45 features, 4 attack families) under stratified cross-validation. The proposed framework achieves an accuracy of 99.89%, macro-averaged F1-score of 0.9568, and Matthews Correlation Coefficient (MCC) of 0.9869. These results show the effectiveness of heterogeneous federated model fusion in privacy-preserving IoMT security.
Predicting stock price direction remains a challenging task due to market volatility, nonlinear cross-asset dependencies, and evolving macroeconomic conditions. Traditional statistical and deep learning approaches have achieved notable progress in modeling temporal dynamics, yet effectively capturing interactions across heterogeneous financial assets remains difficult. Quantum computing offers a way to handle high-dimensional financial data and uncover patterns beyond classical methods. Leveraging this, we introduce the Fusion Quantum Neural Network (FQNN), a hybrid quantum-classical architecture that integrates historical prices with macro-assets as gold price, Bitcoin, and interest rates. Multi-asset features are preprocessed classically, encoded into a parameterized quantum layer, and the resulting measurements are fed into a final classical layer for probabilistic up/down predictions. Experiments on daily market data show FQNN achieves competitive accuracy and greater robustness to abrupt market swings compared to classical benchmarks. Comparative analysis highlights the contribution of both multi-asset fusion and the quantum layer to the overall performance. Targeting NISQ devices, FQNN demonstrates a practical route to promising quantum advantage in finance and highlights the potential of quantum-enhanced feature fusion for more reliable short-term trading strategies.
Participatory fog architectures for IoT exploit underutilized resources to reduce latency, cost, and energy consumption. However, orchestration often relies on a centralized Main Fog Node (MFN) for gateway registration and request assignment, creating a Single Point of Failure (SPOF). When the MFN crashes or becomes overloaded, the control plane is disrupted and request steering may fail, an issue amplified by the high churn of participatory gateways. Existing fault-tolerance solutions frequently require costly coordination or dedicated infrastructure, which is ill-suited to dynamic, resource-constrained green deployments. We propose Virtual-Fog, a lightweight failover mechanism where a small set of guardian gateways monitors MFN liveness, performs a validated takeover, and restores the control state upon recovery. Simulations show that an MFN outage reduces the baseline end-to-end success ratio to 58.7%, while Virtual-Fog achieves 92.8% during the outage and restores overall performance to 97.0%.
The emergence of Large Language Models (LLMs) has revolutionized natural language processing, yet adapting these models to specialized domains via distributed clients is hindered by strict data privacy regulations and constrained communication resources. Although federated learning mitigates the data silo issue by enabling collaborative training without exposing raw local data, federated LLM fine-tuning under heterogeneous multi-task environments still faces critical challenges: i) the massive communication overhead required for synchronizing vast LLM parameters, and ii) severe representation conflicts across divergent tasks, where standard global aggregation often leads to catastrophic performance degradation and structural output collapse in format-sensitive tasks (e.g., Named Entity Recognition). To address these challenges, we propose Per-FedLoRA, a personalized federated fine-tuning framework for multi-task distributed LLMs. First, we adopt a parameter-efficient federated strategy that communicates only Low-Rank Adaptation (LoRA) modules while keeping the backbone LLM frozen locally, ensuring high communication efficiency and data security. Second, we introduce a hierarchical representation decoupling mechanism. Instead of applying unified global aggregation, this mechanism selectively aggregates lower-level adapters to capture universal semantic features while strictly preserving high-level task-specific adapters locally to maintain domain specialization. Third, we formalize this framework to tackle realistic cross-task heterogeneity, effectively mitigating the negative transfer and "weight hedging" effects caused by conflicting optimization spaces. Finally, extensive experimental evaluations demonstrate that our proposed Per-FedLoRA successfully rescues format-sensitive tasks from aggregation collapse, achieving superior multi-task performance and enhanced personalization capability compared to standard federated baselines.
The rapid advancement of embodied intelligence is transforming companion robotics, enabling robots to perceive and interact within their environments. This evolution relies on multi-link wireless connectivity to execute real-time tasks. However, achieving efficient coordination across heterogeneous interfaces is difficult, particularly when balancing low latency, reliability, and energy constraints. To address this, we propose Tacos, a novel Task-Aware Communication Scheduling framework. Tacos integrates Lyapunov optimization with task-aware job abstractions to jointly coordinate communication and computation resources. The algorithm minimizes long-term energy consumption while ensuring queue stability and prioritizing urgent tasks. Extensive simulations show that Tacos significantly outperforms conventional task-agnostic methods in both responsiveness and energy efficiency.
This paper explores the application of THz resonant dipole antennas constructed using composite conductors bundle that integrate carbon nanotubes (CNTs) with copper (Cu). Carbon nanotube conductors demonstrate significant frequency-dependent behavior characterized by complex conductivity. This phenomenon can be effectively modeled using Kubo-type formulations, which elucidate the substantial impedance per unit length observed in THz applications involving various practical arrangements of CNT bundles. Copper is incorporated to alleviate substantial ohmic losses and to stabilize the input impedance in proximity to the resonance frequency. A rigorous electromagnetic model has been formulated in which the composite bundle is characterized by an effective internal impedance per unit length. This impedance is derived from the parallel combination of the conduction pathways of CNT and Cu. The distribution of antenna current and the input impedance are determined by utilizing a thin wire integral equation formulation derived from Hallen's equation, which is solved through the method of moments. The findings underscore the significant influence of dispersion and resistive losses on the conductivity of CNT concerning resonance shifting and impedance characteristics. Additional comparisons with the two limiting cases, namely CNT-only and Cu-only bundles, show that the proposed hybrid CNT-Cu architecture exhibits an intermediate behavior between the two extremes. Furthermore, the results demonstrate that the incorporation of Cu can enhance the practical impedance and improve the overall resonance behavior in THz resonant applications.
This paper proposes a scalable end-to-end scalable multi-layer semantic communication system for efficient multi-user visual tasks in dynamic wireless environments. The transmitter employs a novel hierarchical residual semantic encoder that progressively extracts semantic information through a base layer and multiple enhancement layers. Unlike conventional methods that transmit complete features, it transmits only the base layer features and the residuals between actual features and predictions derived from lower layers, substantially reducing transmission redundancy while preserving task-critical information. At each receiver, a layer-wise semantic fusion strategy combines correctly received consecutive layered semantic features, followed by a differentiated decoding strategy that accommodates both machine users (direct semantic execution) and human users (image reconstruction). Furthermore, we develop a Proximal Policy Optimization (PPO)-based framework that jointly optimizes semantic coding parameters, bandwidth allocation, power allocation, and modulation orders according to time-varying channel conditions. Extensive simulations demonstrate that our approach consistently outperforms conventional Deep Joint Source-Channel Coding (DJSCC) methods across diverse channel conditions.
Cognitive diagnosis, which assesses students’ mastery of knowledge concepts by analyzing their answer records, is one of the core challenges in intelligent education scenarios. In recent years, Graph Neural Network (GNN)-based models have significantly improved diagnostic accuracy. However, existing graph models primarily employ a single graph convolution paradigm, making it difficult to balance global topological structure and local neighborhood information. To address this issue, we propose a diagnostic cognitive framework based on dual-channel graph convolution (DCGCD). Specifically, we first construct a heterogeneous graph among different entities. Subsequently, in the global channel, we employ spectral domain graph convolution to extract overall structural features. In the local channel, we utilize spatial domain graph sampling to sample and aggregate adjacent nodes, efficiently extracting local semantic information. Finally, a linear fusion module obtains complementary and enhanced information. Experimental results on three real-world benchmark datasets demonstrate that DCGCD not only effectively alleviates the limitations of single-graph networks but also outperforms existing mainstream methods in terms of diagnostic accuracy and training efficiency.
Neural Architecture Search (NAS) based on Reinforcement Learning (RL) often suffers from high reward variance and unstable convergence, especially when exploring large design spaces. This paper introduces a hybrid framework that integrates Knowledge Distillation (KD) into Transformer-based RL to stabilize policy optimization and accelerate search efficiency. The proposed KD-TRL-HPO approach leverages a teacher model to guide the student policy through softened signals, reducing exploration noise while preserving architectural diversity. Experiments on the MNIST benchmark demonstrate improved reward stability and superior generalization compared to CE-only TRL-HPO, with the same episode budget and no changes to the search space or controller architecture. By coupling KD with RL, this work offers a scalable and stable solution for neural architecture optimization under limited computational budgets, paving the way for more efficient automated deep model design.
This paper systematically investigates cooperative optimization methods for particle filter algorithms, improved particle flow algorithms, and belief propagation (BP) algorithms under the theoretical framework of nonlinear Bayesian estimation, taking a nine-dimensional state space as the research object to enhance positioning accuracy and real-time performance in dynamic scenarios. Theoretical analysis and simulation results demonstrate that: In comparison to the high particle demand of the particle filter for accurate UAV node positioning, the particle flow algorithm significantly reduces computational complexity through deterministic evolution mechanisms, decreasing particle numbers by 98% while enhancing computational efficiency by 38.9%; The hybrid architecture combining multi-round BP with particle filters can further improve positioning accuracy. The particle flow and BP algorithms effectively enhance computational efficiency and localization accuracy, providing an algorithmic reference for optimizing cooperative localization in UAV swarms.
The rapid growth of the Internet of Things (IoT) deployments has resulted in a massive volume of data generated by smart devices equipped with sensors. The data is typically offloaded to cloud services for convenient and efficient sharing among multiple receivers. However, such cloud-provided encryption mechanisms impose computational burdens on resource-limited IoT hardware and raise data integrity concerns among users. Searchable public key encryption (SPKE) primitive mitigates these issues by ensuring confidential and secure data retrieval. However, existing SPKE approaches in the literature include several critical limitations, including reliance on heavy pairing operations and key-generation authorities, lack of forward secrecy preservation, and support for verifiable ciphertext retrieval in third-party cloud environments. Addressing these concerns, we propose an effective certificateless searchable encryption scheme utilizing standard elliptic curve primitives that eliminates certificate management and key-escrow concerns while avoiding costly pairings. Our construction effectively mitigates information leakage from search queries and access patterns, while incorporating forward-secure verifiable ciphertext retrieval. Additionally, it achieves crucial security guarantees, including ciphertext and trapdoor indistinguishability under adaptive chosen-keyword attacks while also resisting major adversarial threats. Furthermore, comprehensive theoretical and performance evaluations demonstrate the proposed scheme’s efficiency and functional superiority over existing approaches.
With the growing demand for highly reliable and resource-efficient multicast services in next-generation networks, deploying multicast service function chains (SFCs) poses significant challenges due to complex interdependencies among VNF placement, routing, and link protection. Unlike traditional unicast scenarios, multicast SFCs must simultaneously serve multiple destinations while ensuring reliability and minimizing resource overhead. This paper presents a reliability-aware multicast SFC deployment framework that integrates centralized optimization with distributed decision-making. A unified reliability-constrained model is formulated to jointly determine VNF placement, multicast routing, and link backup configuration. To enable tractable optimization, the inherently nonlinear reliability-constrained formulation is transformed into an equivalent linear model, achieving coordinated reliability assurance and resource efficiency.To enhance scalability, a game-theoretic distributed algorithm is designed to model domain interactions and iteratively converge to near-optimal solutions with lower complexity. Simulation results show that the proposed framework achieves an effective reliability–cost trade-off, reducing resource consumption by about 15–23% compared with heuristic schemes. The proposed approach provides a flexible and scalable solution for reliable multicast SFC deployment in 6G-oriented networks.
Accurate indoor localization (IL) in 6G millimeter-Wave (mm-Wave) networks is a critical enabler for autonomous robotics, asset tracking, and Internet of Things (IoT) sensing, yet existing methods face a persistent trade-off between accuracy, robustness, and deployability: received signal strength indicator (RSSI)-based approaches lack spatial resolution; Deep Learning (DL)-based methods require large labeled datasets and offer limited interpretability; and the multipath delay profile (MDP) fingerprinting baseline Multipath Component Analysis (MCA) suffers from non-bijective multipath component (MPC) matching and environment-specific parameter tuning. This paper introduces MPUrge-MAP, a white-box, training-free MDP fingerprinting-based IL method that adapts the MPUrge tri-phase iterative MPC matching algorithm to direct fingerprint-based positioning, to the best of our knowledge, its first application to this task. Two targeted improvements address the core limitations of the original MPUrge formulation: a normalized pattern-based dissimilarity metric that restores magnitude balance between the distance-based and pattern-based components, and an enhanced overall dissimilarity score that penalizes unreliable reference point (RP) candidates by accounting for match coverage and vector length. Together, they yield an efficient, parameter-robust pipeline requiring no floor plan, no training data, and no dense calibration grid. Simulations on the validated NIST quasi-deterministic (QD) 60 GHz mm-Wave channel simulator show that MPUrge-MAP achieves a mean localization error of 0.330 m corresponding to a 54% reduction over MCA (0.717 m) and 13.8% over the original MPUrge (0.383 m). In addition, MPUrge-MAP shows a competitive robustness against AWGN noise and environmental obstruction across all tested configurations.
This study presents a novel approach to resource allocation in edge-driven IoT networks using a classification-based neural network. IoT devices face significant constraints due to limited resources and increasing user demands. While edge computing offers a promising solution by processing data closer to the source, it introduces challenges such as dynamic network conditions and infrastructure limitations. To address these challenges, we propose a framework based on a classification-based neural network that jointly evaluates available edge resources and selects the most suitable one for task execution. The model, implemented using TensorFlow, performs holistic resource comparison, enabling fast task allocation decisions under fluctuating conditions. Experimental results show that the proposed approach achieves a task completion rate of 56% within priority constraints in a dynamic testing environment, demonstrating stable behaviour under resource volatility.
