ABSTRACT The evolution of sixth‐generation (6G) wireless networks demands ultra‐reliable low‐latency communication (URLLC), massive connectivity, and high‐capacity data transmission in highly dynamic environments. Space–Air–Ground Integrated Networks (SAGINs) have emerged as a promising architecture by seamlessly integrating satellites, unmanned aerial vehicles (UAVs), and terrestrial infrastructure to provide ubiquitous connectivity. However, stochastic traffic arrivals, UAV mobility, time‐varying wireless channels, and the coexistence of enhanced Mobile Broadband (eMBB) and URLLC services make traffic offloading and resource allocation highly challenging. These factors transform the optimization task into a stochastic mixed‐integer nonlinear programming (MINLP) problem. To address this challenge, this paper proposes a Quantum Federated Reinforcement Learning (QFRL)‐based traffic offloading framework for RSMA‐enabled SAGINs. The optimization problem is formulated as a constrained Markov decision process (CMDP), allowing distributed small cells to jointly optimize traffic offloading ratios, bandwidth allocation, RSMA power distribution, and UAV trajectory planning while satisfying stringent delay and reliability requirements. A variational quantum circuit (VQC)‐based actor‐critic architecture is developed to improve learning efficiency and policy representation in high‐dimensional continuous action spaces. In addition, a federated aggregation mechanism enables privacy‐preserving distributed learning and scalable coordination across the space, air, and ground segments. The proposed framework employs temporal‐difference learning and parameter‐shift gradient optimization to ensure stable convergence under stochastic network dynamics. Simulation results demonstrated that the proposed QFRL framework reduces traffic dropping probability by 28%–35%, decreases URLLC delay by 22%–30%, improves network availability by 18%–25%, and enhances traffic offloading efficiency by 20%–27% compared with Differentiated Federated Soft Actor‐Critic (DFSAC), Double Q‐Learning delay sensitive replay memory (DSRPM), and Nash Equilibrium Iteration Offloading (NEIO‐G) schemes, respectively.
ABSTRACT The rapid growth of heterogeneous network environments such as the Internet of Things (IoT), Industrial IoT (IIoT), cloud computing, and software‐defined networks has significantly increased exposure to sophisticated cyberattacks, making intrusion detection a critical component of modern cybersecurity infrastructures. Traditional intrusion detection systems and conventional machine learning techniques often face limitations when handling high‐dimensional network traffic, class imbalance, and evolving attack patterns, resulting in reduced detection performance and limited scalability under complex network environments. These challenges reduce their effectiveness in practical, large‐scale deployments. To overcome these issues, this paper proposes a hybrid intrusion detection framework based on an Autoencoder and a TabTransformer, optimized using the Whale Optimization Algorithm (WOA). The Autoencoder is employed to perform unsupervised feature learning, transforming high‐dimensional network traffic data into compact and noise‐resistant latent representations. These latent features are then processed by the TabTransformer, which utilizes multi‐head self‐attention to capture complex inter‐feature relationships and enhance classification performance. The WOA is incorporated to automatically optimize key hyperparameters, improving convergence speed, stability, and generalization capability of the model. The proposed framework is primarily evaluated using the CIC‐IDS2018 benchmark dataset. In addition, supplementary cross‐dataset validation on the CIC‐IDS2017 and UNSW‐NB15 datasets is conducted to assess the generalization capability of the proposed framework. Experimental results demonstrate that the proposed model achieves an accuracy of 99.87%, precision of 99.85%, recall of 99.88%, and an F1‐score of 99.86% while maintaining very low false alarm and false negative rates. Comparative analysis with existing deep learning‐based intrusion detection approaches confirms the superior and balanced performance of the proposed method. Overall, the Hybrid Autoencoder–TabTransformer framework provides an effective intrusion detection solution that demonstrates strong performance under the evaluated experimental conditions.
ABSTRACT The convergence of 6G communications and intelligent photovoltaic (PV) grids introduces severe cybersecurity risks, including sophisticated DDoS attacks, data manipulation, and unauthorized access, which conventional intrusion detection systems fail to adequately address in such dynamic and data‐intensive environments. This paper proposes a novel AI‐based intrusion detection framework tailored for 6G‐connected PV smart grids, with two core innovative modules: Multi‐modal Adaptive Variational Autoencoder (MA‐VAE) and Hierarchical Spatiotemporal Attention Network (HSTA‐Net). MA‐VAE features modality‐specific encoders, a shared latent space with disentangled representation learning and an adaptive weighting mechanism for dynamic adjustment of multi‐modal contribution weights in heterogeneous data fusion; HSTA‐Net integrates multi‐scale temporal convolution and gated attention, with a spatiotemporal attention mechanism focusing on key time points and feature dimensions for cross‐scale spatiotemporal dependency modeling. Experimental evaluations demonstrate that the proposed framework outperforms state‐of‐the‐art baseline methods in terms of detection accuracy and computational efficiency, achieving a 30% reduction in computational cost compared to recent large language model (LLM)‐based intrusion detection models. This work delivers a scalable, low‐latency security solution for 6G‐integrated PV grids and advances adaptive intrusion detection through effective multi‐source fusion and temporal modeling.
ABSTRACT With the growing adoption of electric vehicle charging systems (EVCS), energy shortages and power performance challenges have become significant concerns. This study developed an intelligent energy management system (EMS) integrated with a photovoltaic (PV) system and utility grid to enhance the operational efficiency and sustainability of EVCS. The proposed system utilizes a novel optimization technique, the Markov decision‐based snake optimization with maximum power point tracking (MDbSO‐MPPT), to effectively manage power and energy performance. The EMS optimizes energy distribution, minimizes grid dependency, and ensures cost‐effective and reliable operation of the charging station. Simulation results demonstrate that the developed system achieves a remarkable 99.8% efficiency, outperforming conventional models with 98% and 99.5% efficiencies. The grid power consumption was also reduced to 63 kW compared to 100 and 182 kW for the benchmarks. The system also optimizes the PV power utilization and battery state of charge, reducing unnecessary grid interactions and improving sustainability. By leveraging renewable Energy and advanced optimization strategies, this intelligent EMS provides a scalable and efficient solution for modern EV charging needs.
ABSTRACT Intelligent reflecting surfaces (IRS) have emerged as a transformative technology for B5G/6G networks, offering the ability to actively reshape the wireless propagation environment and support reliable line‐of‐sight communication. In the context of intelligent transportation systems, however, meeting the stringent reliability and latency demands of ultra‐reliable low latency communication (URLLC) remains a significant challenge, particularly under packet collision scenarios. To address this, we propose a retransmission‐based protocol called retransmission mechanism with reserved bandwidth in IRS (RMRI), tailored to satisfy URLLC requirements in IRS‐assisted vehicular networks. RMRI leverages a finite set of reserved bandwidth channels to instantly retransmit collided packets, combined with a packet diversity principle using 3‐, 5‐, and 7‐packet duplication and a slotted Aloha mechanism operating within 0.2 ms time slots. We evaluate the protocol in terms of the minimum subcarriers needed to achieve 99.9999% reliability within a 0.2 ms latency threshold, maximum supported users, end‐to‐end latency, and path loss using the 3GPP UMi model. Results show that RMRI consistently outperforms conventional schemes, offering reduced subcarrier usage, enhanced reliability, and low latency, effectively meeting the demands of next‐generation intelligent transportation systems.
ABSTRACT Quantum Computing (QC) has emerged as a promising paradigm capable of solving complex computational problems more efficiently than classical approaches. Quantum‐inspired algorithms provide highly efficient, optimized solutions that can revolutionize various industries by enabling faster data processing, secure communications, and innovative advancements. In cloud security, Machine Learning (ML) techniques play an essential role in detecting and mitigating cyber threats. The integration of QC with ML, known as Quantum Machine Learning (QML), introduces novel security solutions that enhance the resilience of cloud infrastructures against cyberattacks. Moreover, incorporating Generative AI into QML frameworks can further bolster cloud security through advanced anomaly detection, automated response generation, and the synthesis of secure synthetic data for training models. Additionally, in IoT‐enabled sustainable smart cities, information fusion techniques (including multi‐sensor and multi‐process fusion) are essential for ensuring robust cloud security, which is paramount to protecting interconnected systems, data privacy, and critical infrastructure. This review article provides a comprehensive survey of QML‐based approaches, augmented by generative AI and information fusion, for enhancing cloud security in these environments. We review over 150 research articles, examining the key principles of QML, its advantages over classical ML techniques, and its applications in securing cloud environments. Furthermore, the paper discusses the challenges and opportunities in implementing QML, generative AI, and information fusion for cloud security within smart city ecosystems. By offering a detailed analysis, this study aims to contribute to the advancement of quantum‐driven security frameworks for next‐generation smart cities.
ABSTRACT As libraries adapt to the digital age and the challenges posed by the evolving information landscape, adopting emerging technologies becomes paramount. With its reputation for security, transparency, and decentralization, blockchain technology has come to light as a potentially revolutionary instrument that might completely change library services and operations. This review paper explores how blockchain has been applied and adopted to suit libraries, shedding light on its transformative capabilities and its benefits. It begins with an introduction to blockchain technology, characterized by critical features and historical context. It then delves into specific use cases within libraries, ranging from cataloguing and metadata management to digital asset preservation and copyright management. Real‐world case studies and examples are presented to illustrate the practical implementation of blockchain in libraries and their consortia. While blockchain technology holds great potential, this article also identifies the implementation obstacles that libraries may encounter, such as budgetary constraints, issues with scalability, and compliance with regulations. Ethical and privacy considerations are examined, emphasizing the need to protect patron data and ensure responsible usage of blockchain technology in a library environment. Additionally, this research review outlines the many advantages and possibilities that blockchain offers libraries, including increased user trust, decreased fraud, and flexibility to accommodate changing user requirements.
ABSTRACT Sequential federated learning (SFL) enables collaborative model training across clients in a chain manner, providing communication‐efficient benefits over traditional all‐gather parallel federated learning (PFL). However, SFL training often suffers from slow convergence and performance degradation due to nonidentically distributed (non‐IID) data distribution. In motivation experiment, we find that model decoupling by partitioning the model into shared and personalized parameters and using just a few personalized parameters with large gradients can improve SFL training performance. Based on above findings, we propose SeqFedRPC , a novel model decoupling based SFL framework with regularized parameter clustering. We introduce a regularization term to promote parameter sparsification and amplify gradient differences, which aids in gradient‐based parameter clustering. Then, we employ a clustering‐based scheme to adaptively decouple the model parameters into shared and personalized subsets, thereby addressing the challenge of non‐IID data by adapting global knowledge with shared parameters and client‐specific distributions with personalized parameters. Extensive experiments on eight benchmark datasets demonstrate that SeqFedRPC surpasses eight SOTA methods, with each client personalizing less than 10% of the total parameters on all datasets at .
ABSTRACT While encrypted traffic ensures secure communications, it also provides covert channels for some advanced attacks, making them difficult to detect. Detection is particularly difficult in the context of intelligent connected vehicles (ICVs), where there are two huge challenges: (1) ICVs dynamic network topologies induce traffic pattern drift, rendering traditional static detection models obsolete. (2) The coexistence of heterogeneous sensors and diverse data sources blurs traffic boundaries, degrading detection efficacy and manifesting persistent false negatives. To address these challenges, we propose an architecturally novel paradigm GraphSAGE‐KAN. Initially, time‐aware flow graphs are constructed through the encoding of communication mechanisms into dynamically weighted edges. This approach can bridge transport‐application semantics and mitigate the problem of blurred boundaries between traffic. Subsequently, node features are optimized through Laplacian scores and Spearman's rank coefficient analysis, which can address the problem of traffic pattern drift, ensuring the selection of discriminative features. Finally, graphs are processed using a GraphSAGE‐KAN, which effectively “lifts” the aggregated features into a higher‐dimensional manifold, enabling the network to approximate any continuous function over its local receptive field. Experiments were conducted using the IoT environment datasets (public datasets) and real‐world attack datasets. Compared with the current method, the proposed method shows better performance in terms of accuracy, precision, recall, F1 score, and false positive rate.
ABSTRACT Design for Test (DFT) enhances the reliability of integrated circuits (ICs). However, the use of scan‐based methodologies exposes internal nodes within ICs to possible side‐channel attack vectors. This paper proposes a novel secure architectural framework around the DFT logic to protect against SCA and other forms of confidential data leakage from the chip. This novel architecture implemented is a Low‐Power Q‐gate Secured Scan Register (LP‐Q‐SSR), that bridges the gap of traditional scan‐based designs that reduces the toggling activity, power usage and secures the chip. The proposed design incorporates improved low‐power design with Linear Feedback Shift Register (LFSR) driven mux flop which is a counterpart of conventional scan methodology incorporating the Q‐gating and X‐filling techniques to maintain the design test, power and security conscious. It uses Multi‐Polynomial Asymmetric Dual (MP‐Asym Dual) LFSR and its tapping mechanism to control the scan data flow and uses test strategies like Q‐gating technique and X‐filling algorithm for optimizing the power. This makes the hackers difficult to encrypt the test‐mode and to fetch the confidential data of an IP or IC or SoC. The proposed secured scan architecture demonstrates significant improvements over the conventional scan‐based architecture, achieving ∼1.81× reduction in test pattern count, ∼5.45× faster runtime, ∼27.13% lower switching activity, and ∼50.9% power savings while supporting a 128‐bit design. Additionally, the architecture enhances testability (∼1.5%) and memory efficiency (∼1.25×) with minimal area overhead of only ∼2.17%, highlighting its effectiveness for low‐power and secure VLSI testing.
The Internet of Things (IoT) and its applications occupy a prominent place in contemporary research. The IoT, with its inherently heterogeneous behavior offers solutions to numerous problems and has become inseparable from human life. In this paper, an Optimized Multi-Channel Graph Neural Network Framework for Intrusion Detection in Internet-of-Things Environments (DGSEMGNN-IDF-IoT) is proposed. Here, the data collected from the Telemetry of Network-Internet of Things (ToN-IoT) dataset are used. To execute this, the input data are given into the pre-processing stage using the Dual Central Difference Kalman Filter (DCDKF) to eliminate redundant values and replace missing ones. The pre-processed data are supplied to the feature selection stage using the Ebola Optimization Search Algorithm (EOSA) to select the optimum features. The selected features are fed into the Dynamic Global Structure Enhanced Multi-Channel Graph Neural Network (DGSEMGNN) to effectively categorize the data as normal, backdoor, scanning, password, Cross-Site Scripting (XSS), ransomware, injection, Denial of Service (DOS), Distributed Denial of Service (DDOS) and Man-in-the-Middle (MITM) attack. Then, the Wolf-Bird Optimizer (WBO) is used to enhance the weight parameters of DGSEMGNN. The DGSEMGNN-IDF-IoT is implemented, and the performance metrics, such as accuracy, f1-score, specificity, sensitivity, Receiver Operating Characteristic (ROC), error rate, and computation time, are analyzed. The DGSEMGNN-IDF-IoT approach attains 6.80%, 10.41%, and 6.60% higher accuracy, 6.95%, 8.02%, and 10.9% higher f1-score when compared with existing methods.
Space-air-ground integrated networks (SAGIN) provide ubiquitous connectivity for large-scale internet of things (IoT) applications by integrating satellite, aerial, and terrestrial networks. However, secure and efficient data authentication in SAGIN remains challenging due to network heterogeneity, dynamic topology, and resource-constrained devices. Traditional public key infrastructure based (PKI-based) schemes incur high certificate management overhead, while identity-based approaches suffer from key escrow, limiting their applicability in SAGIN environments. This paper proposes a hierarchical lightweight certificateless aggregate signature scheme for secure data authentication in SAGIN. In the proposed scheme, IoT devices generate lightweight certificateless signatures, which are hierarchically aggregated at unmanned aerial vehicles and high-altitude platforms without performing pairing operations. The final aggregated signature is verified at powerful satellite or ground control stations, significantly reducing computation and communication overhead. Security analysis shows that the scheme is existentially unforgeable under adaptive chosen-message attacks and resistant to pollution attacks. Performance evaluation through theoretical analysis and simulation demonstrates low overhead, constant-size aggregated signatures, and good scalability, making the proposed scheme suitable for large-scale SAGIN-enabled IoT systems.
Basketball action recognition based on wearable devices and deep learning has been widely applied in the field of behavior analysis. However, there is an urgent need for real-time recognition of athlete movements, precise technical analysis, and data privacy protection in modern basketball training and competitions. To this end, this article proposes a low-latency basketball action recognition and game technology analysis framework based on multimodal spiking neural networks (SNNs) under the space air ground-integrated network (SAGIN) environment. This study utilizes multiple wearable sensors to synchronously collect high-frequency multimodal time series signals. To effectively process these sensing data with high spatiotemporal complexity, we design an effective feature extraction backbone based on SNNs which can capture temporal information of time series data with low computational complexity and low latency. In addition, to improve the recognition accuracy and robustness of complex basketball actions, we exploit the joint learning module to enhance multimodal information fusion. At last, to address the issue of data silos and ensure user data privacy, we adopt a personalized federated learning paradigm, allowing all parties involved to collaboratively optimize the global model without sharing local raw data. Our proposed method has been evaluated on various public datasets, including OPPORTUNITY, PAMAP2, SKODA, and our self-built basketball action dataset. The extensive experimental results demonstrate that our approach achieves significant performance and is more robust to noise compared to other deep models. Our proposed framework can achieve high-precision real-time action classification and fine-grained technical statistical analysis while ensuring low communication and computing overhead, providing an effective technical solution for the intelligent and personalized development of basketball sports in the context of integrated aerospace networks.
Traffic accident prediction using dashboard cameras is of great importance in the safety of self-driving systems and the mitigation of accidents. This is because the prediction of accidents early is not easy because of the complex traffic conditions and the large variety of different object movements. In order to solve this problem, a new model of predicting traffic accidents will be suggested based on the Dual-Path transformer fusion (DualTF) model. Video itself is divided into time and space. There is one frame that contains the background and position of the object information, which is spatial information. The camera and object movement through various frames is what gives the time information. The proposed DualTF model will detect the co-occurring objects on the road using the Unified Transformer Framework. The Optical Flow is obtained by using the FlowFormer. Subsequently, the space and time characteristics will be obtained from co-occurring objects and optical flow. The characteristics will then be amalgamated to predict traffic accidents. Lastly, the proposed approach has a greater degree of performance compared to other approaches and achieves an accuracy of 99.5%.
The integration of vehicular ad hoc networks (VANETs) with human activity monitoring networks (HAMNs) is advancing intelligent mobility and connected healthcare. However, achieving accurate anomaly detection while maintaining sustainable electric power utilization remains challenging due to dynamic vehicular mobility and fluctuating energy demands. Existing approaches such as VANET security awareness framework (VSAF), secure passenger health assessment using vehicular communications (SPHA-VC), and bio-inspired VANET routing (BIVR) mainly focus on security awareness, passenger health monitoring, or routing optimization, but they lack adaptive mechanisms for joint anomaly detection and energy-efficient network management in highly dynamic environments. To address these limitations, this paper proposes ECO-SHIELD, a hybrid energy-conscious self-healing intelligent learning and detection framework that integrates graph neural networks (GNNs) with a modified particle swarm optimization (M-PSO) algorithm. The GNN module models spatial-temporal interactions among vehicular and human activity nodes, while M-PSO dynamically optimizes network parameters and energy thresholds to improve detection reliability and power efficiency. The comparative analysis reports detection accuracy values of 86.3%, 88.1%, and 89.4% for VSAF, SPHA-VC, and BIVR, respectively, in comparison with 98.4% achieved by ECO-SHIELD. Similarly, response latency values of 168, 149, and 134 ms and energy consumption values of 345, 318, and 296 mAh have been included for the baseline methods, which clearly highlight the improvements obtained by the proposed framework. In addition, the comparative results now present packet delivery ratios of 91.2%, 92.8%, and 94.1% for the baseline methods against 97% for ECO-SHIELD.
In intelligent transportation systems, Vehicular Ad Hoc Networks (VANETs) provide real-time vehicle-roadside infrastructure communication to improve traffic safety and efficiency. However, VANETs are vulnerable to sophisticated network assaults like DoS, Sybil attacks, and message tampering due to their mobility, lack of centralization, and changeable topology. Traditional IDSs fail to detect such threats without sacrificing vehicle privacy. A novel IDS system, the Distributed Adaptive Network Anomaly Guard (DynaGuard), leverages distributed machine learning via privacy-preserving federated learning to address these concerns. DynaGuard lets cars train local models using network traffic patterns, detect abnormalities in real time, and adjust threat levels using ensemble-based learning. To conclude, DynaGuard protects vehicle networks from emerging cyber threats and enables intelligent transportation systems with secure, dependable communications. In a simulated VANET dataset, DynaGuard outperforms typical IDS techniques with a 97.1% detection rate, 19% fewer false positives, and low communication cost.
We propose a Beidou-augmented Sparse Variational GNSS/INS Combined System (BSV-GIIS) to tackle the difficulties of high-accuracy positioning in intelligent sugarcane agriculture, where thick vegetation and multipath interference reduce the effectiveness of traditional tightly coupled Kalman filter (TCKF) approaches. The system introduces a hierarchical sparse approximate message passing (HSAMP) framework, which reformulates the fusion problem into a two-layer variational inference process with adaptive sparsity constraints, thereby capturing the inherent noise sparsity in agricultural environments. At the global scale, a mixture of sparse Gaussians approximates the posterior state distribution, whereas the local scale improves estimates by modeling Beidou-specific residuals with heavy-tailed distributions. In addition, a compact hybrid Kalman-variational filter (LHKVF) processes inertial data by means of selective Mahalanobis gating, which permits effective outlier rejection. The 5 cm accuracy is based on median performance at nominal conditions, while RMS errors in dense canopy environments are above 8 cm. This shows that although the median error values are within the 5 cm range for open and partial canopies, the RMS errors increase to 8.3 cm in dense canopies, thus proving the limitation of sparse assumptions in dense foliage occlusions. This shows that although the system attains centimeter-level accuracy in nominal situations, there is a performance degradation in dense foliage occlusions. Implemented on a Xilinx Versal ACAP, BSV-GIIS shows effective handling of sparse matrix computations and precise coordination between IMU and GNSS data flows. Experimental outcomes establish its advantage in accuracy and processing speed, which renders it a feasible option for embedded systems with limited resources in precision agriculture.
Data security in Electronic health records (EHRs) plays a crucial role in the healthcare system because the EHRs contain sensitive information about patients' personal and medical data, making them a prime target for security breaches and unauthorized access. EHRs are often accessed and updated by many healthcare professionals, so robust data security measures are necessary to control and monitor access to patient records. To safeguard patient data against unauthorized access, numerous deep learning (DL) approaches have been introduced. However, many of these methods face significant limitations, such as low detection precision, excessive computational demands, and susceptibility to overfitting. In response to these challenges, this study introduces a novel hybrid model that combines Recurrent Backpropagation Learning (RBPL) with a Bi-directional Long-Term Deep Belief Network (BLDBN). This integrated approach aims to strengthen the security of Electronic Health Records (EHRs) within IoT-enabled cloud infrastructures. The model processes input data sourced from both the Healthcare Dataset and the IoT Healthcare Security Dataset, incorporating preprocessing techniques to improve data quality. The RBPL-BLDBN utilizes a feature selection module for selecting the more significant features. The selected features improve the model's efficiency by reducing computational complexity and improving accuracy. The RBPL-BLDBN technique designs an RBPL approach to handle time-dependent data, leveraging its recurrent nature to track sequential health records and detect anomalies or unusual behavior patterns. Furthermore, the BLDBN component captures hierarchical features across multiple layers and enhances the model's ability to detect intricate patterns in healthcare data. The proposed RBPL-BLDBN technique not only enhances the security of healthcare data but also detects and prevents unauthorized access or anomalies in real time. Extensive evaluation of the RBPL-BLDBN framework across multiple performance indicators reveals its superior capability in healthcare data protection. The model achieves impressive scores, attaining 98.84% in accuracy, 98.55% in precision, and 97.87% in specificity, which highlights its efficiency and robustness. The comparative analysis further demonstrates that the RBPL-BLDBN not only conserves computational resources but also surpasses the performance of several existing security models in the domain.
In dynamic vehicular ad hoc networks (VANETs), high vehicle mobility and time varying communication links make task offloading highly uncertain. As the numbers of vehicles and tasks increase, the joint decision space grows rapidly under this uncertainty, which further complicates offloading optimization. Traditional reinforcement learning methods therefore often suffer from action-space explosion and unstable convergence. To address this issue, we propose a hierarchical task-enhanced fully decentralized multi-agent proximal policy optimization method (HT-FMPPO), which integrates a hierarchical reinforcement scheduling framework (HRS) and a target entropy driven adaptive entropy regularization mechanism (TEAR). The HRS decomposes joint offloading into batch level quota planning and task level execution location selection, thereby reducing decision complexity in large action spaces. The TEAR adjusts the entropy coefficient online based on the gap between current policy entropy and target policy entropy, thereby improving exploration in early training and reducing excessive randomness during convergence. The experimental results show that HT-FMPPO reduces request completion time and total energy consumption across different vehicle and task scales, while achieving better convergence and stability.
Position integrity is a critical requirement for the safe operation of vehicles in Intelligent Transportation Systems (ITS). Manipulated position messages lead to unsafe decisions and traffic confusion and create possible accidents. Existing approaches primarily focus on improving localisation accuracy and securing vehicle identities, rather than detecting false positional behaviors. Also, real-world communication challenges make integrity verification more difficult in dense vehicular environments. To address these challenges, we propose a Cascaded-Deep Ensemble Learning (CADE) model to assess positional integrity in VANET environments. The CADE model evaluates the trust of shared position information using a multi-stage pipeline. It combines motion-prediction-based Position Error (PE) estimation, adaptive safety-bound evaluation via Adaptive Dynamic Level (ADL), temporal-behavior verification using Temporal Consistency Score (TCS), and cooperative validation via Cooperative Plausibility Index (CPI). These integrity indicators are fused and evaluated using lightweight ensemble classifiers to obtain the position information of active vehicles. The proposed CADE framework is evaluated on the VeReMi dataset; experimental results show that the approach achieves a high detection accuracy of 99.50% and demonstrates a strong ability to identify false-position attacks even under congested vehicular communication conditions.
