High-precision vehicle trajectory prediction method has become the key technology of automatic driving, traffic management and path planning. To address low trajectory prediction accuracy and limitations in scene adaptability of urban autonomous driving vehicles, a Stacked Long Short-Term Memory Graph Convolutional Network model (STACKING-LSTM-GCN) is proposed. Our method first constructs a multi-feature fusion model to analyze the spatial, temporal, traffic and driving behavior features related to the autonomous driving trajectory. Secondly, GCN is used to construct the vehicle interaction graph by updating the adjacency matrix in real time, so as to accurately capture the spatial interactions among vehicles over time. Finally, based on the rich feature information, the superposition method is used to predict the vehicle trajectory, and the experimental evaluation is carried out by comparing the average RMSE values under different teacher enforcement rates, hidden layer units and batch size settings. In the four autonomous driving road scenarios, the performance is optimal when the teacher forcing rate is 0.5, the hidden size is 128, and the batch size is 64. Experimental tests on the NGSIM, HighD and Argoverse datasets show that our proposed model exhibits better performance than other models in terms of MAE and RMSE, highlighting its effectiveness under various road conditions. Stacking-LSTM-GCN maintains robust prediction performance with limited training data, demonstrating superior data efficiency compared to baseline methods, which further enhances the advantages of this method.
In IoT-enabled healthcare scenarios, patients submit symptoms and physiological parameters to physicians via online consultations for diagnostic and treatment advice. While the Internet improves medical accessibility, it also introduces challenges in the confidentiality protection of user identities and health data. As cross-institutional collaboration becomes increasingly critical for enhancing care quality, the potential for redundant consultation processing also increases; efficient deduplication of redundant consultations is therefore needed to conserve resources. Additionally, since medical team compositions change frequently, dynamically revoking access permissions is essential for maintaining data security. In this paper, we propose a privacy-preserving scheme with dynamic revocation (PPS-DR) for medical consultation services (MCS). Specifically, our scheme achieve the anonymity of users’ identities, confidentiality of the consultation content, and linkability of a user’s multiple consultations. The medical server (MS) performs equality tests to deduplicate consultations, thus enhancing operational efficiency. Furthermore, upon physicians’ departure from the medical team, the MS can instantly revoke their decryption privileges without requiring patients to reprocess the corresponding data. Our scheme offers provable security in the random oracle model (ROM) and delivers significantly greater efficiency than existing schemes. Specifically, its computation and communication costs account for only 86.8% and 75.1% of the most efficient baseline, respectively, making it particularly appropriate for MCS applications.
Shared Energy Storage Centre (SESC) can enhance the flexibility of multi-microgrid (MMG) systems and improve renewable energy utilization. However, SESC planning remains constrained by long-term renewable and load uncertainty, stringent siting requirements, and the coupling between capacity allocation and profitability. This paper proposes an integrated framework combining large language model (LLM)-based multi-agent systems with exploration-enhanced deep reinforcement learning. The LLM agents address renewable, load forecasting and SESC siting, while the reinforcement learning module performs economic optimization. The upstream Multi-ChatGPT-Agent rolling forecast module enables multi-year joint forecasting of wind power, photovoltaic generation, and load. Comparative experiments show improved forecasting accuracy, and the prediction results are converted into representative datasets through clustering for downstream optimization. Guided by these forecasts, the ChatGPT-Agent infers SESC locations from microgrid coordinates and scale metrics while jointly considering wiring-distance minimisation, transmission-loss-cost minimisation, and practical siting constraints expressed in natural language. This allows language-based engineering requirements that are difficult to encode as mathematical or programmable constraints to be incorporated into the siting process. Subsequently, an Adaptive Count-Based Exploration Deep Q-Network (ACBE-DQN) is constructed to jointly learn capacity allocation and profitability strategies. By enhancing exploration through adaptive intrinsic reinforcement, ACBE-DQN mitigates sparse and non-stationary profit signals. Empirical studies show that, compared with standard DQN and Q-learning, the proposed model improves convergence stability and economic efficiency while maintaining feasible energy-state trajectories and operational constraints.
This paper investigates the problem of FDI attack detection and attack compensation for networked control systems under finite-time boundedness framework. In contrast to existing research that primarily focuses on attack detection or addresses the asymptotic stability of systems under cyber attacks, this study considers the limitations of network bandwidth and jointly designs an attack estimation and real-time compensation control framework to establish finite-time boundedness for the system under external disturbances and attacks. Firstly, a dynamic quantizer is designed to adaptively adjust quantization precision based on signal variations, and a conventional residual evaluation mechanism is employed to detect attacks. Secondly, a triggered compensator is designed to mitigate the effects of malicious attacks on the system states. Subsequently, sufficient conditions guaranteeing the finite-time boundedness of the system and satisfying the H∞ performance requirement are formulated in terms of linear matrix inequalities. Finally, the effectiveness of the proposed method is validated through simulation results.
Cloud-based speaker verification systems face critical privacy concerns regarding biometric template leakage, particularly in practical IoT scenarios such as smart access control, connected vehicle authentication, and remote financial account opening, where compromised voiceprints may cause economic losses or enable deepfake-based impersonation attacks. While homomorphic encryption (HE) protects data confidentiality, pure HE schemes are limited when executing nonlinear operations, such as division and square roots, required for high-precision cosine similarity evaluation. Existing methods therefore rely on polynomial approximations that introduce cumulative numerical errors and degrade recognition accuracy, particularly in clustered embedding spaces with extreme decision thresholds. To overcome these limitations, we propose HPSV, which combines fully homomorphic encryption (FHE) with secure two-party computation (2PC). HPSV offloads critical nonlinear computations to lightweight secure 2PC protocols, reducing approximation error while preserving biometric privacy. Experimental evaluations on the LibriSpeech and TIMIT datasets demonstrate that HPSV closely matches the discriminative power of the unencrypted baseline. Threshold analysis further confirms that HPSV maintains robust authentication performance under strict decision thresholds. Compared with approximation-based HE schemes, HPSV provides a concise and accuracy-preserving solution for secure IoT biometric authentication.
Securing high-traffic public venues requires detecting concealed threats without disrupting pedestrian flow or compromising privacy. While ceiling-mounted Ultra-Wideband (UWB) radar offers a promising non-intrusive solution, it faces severe challenges such as signal distortion from the overhead view and dominant floor clutter that masks faint metallic signatures. To address these limitations, we propose SAFE, a lightweight pass-through detection framework utilizing a single UWB sensor. Instead of relying on computationally expensive deep learning models, SAFE adopts a physics-aware approach: it aligns radar signals with the pedestrian’s trajectory to rectify geometric distortions and extracts distinct spatial and kinematic features to isolate metallic objects from body motion. Extensive experiments demonstrate that SAFE achieves robust discriminative performance with an Area Under the Curve (AUC) of 0.921 and over 3,800× greater computational efficiency compared to state-of-the-art models, and a benchmark-based latency projection indicates its potential for real-time, privacy-preserving edge deployment.
Integrated sensing, communication, and computation (ISCC) has emerged as a key research focus for sixth-generation (6G) communication systems, playing a crucial role in advancing intelligent transportation through enhanced connectivity and automation. However, this integration currently lacks systematic theoretical foundations and key technological support, presenting critical challenges in analyzing ISCC’s fundamental performance and balancing tradeoffs among communication, sensing, and computation to resolve inherent conflicts. This paper presents a comprehensive study on ISCC by integrating the mobile edge computing (MEC) paradigm with integrated sensing and communication (ISAC) technology, providing novel insights into performance analysis and design optimization. We focus on scenarios involving MEC-Empowered MIMO-ISAC for vehicular networks, aiming to thoroughly analyze the fundamental performance of ISCC in enhancing vehicle communication, situational awareness, and real-time decision-making. Specifically, we consider both downlink and uplink ISAC scenarios to explore the comprehensive performance of ISCC. For downlink ISAC (D-ISAC), we explore three design paradigms: sensing-centric (S-C), communications-centric (C-C), and Pareto optimal design, which are critical for optimizing vehicle interactions and improving road safety. For uplink ISAC (U-ISAC), we analyze the causality of offloading computing tasks by vehicles, performing a Pareto optimal design under the power constraints at the BS. Through the Pareto optimal design, a three-dimensional CMR-SR-CPR (Communication Rate, Sensing Rate and Computation Rate) rate region is characterized to achieve a communication-sensing-computation performance tradeoff. This approach improves data processing efficiency and provides critical insights for ISCC resource optimization and tradeoff design.
In recent years, Mobile Crowd Sensing (MCS) has emerged as a research hotspot, offering an innovative mode of perception widely applied in areas such as intelligent transportation and environmental monitoring by utilizing mobile devices for data collection and information analysis. However, traditional UAV-assisted mobile crowd sensing struggles to fully meet the diverse objectives of large-scale perception tasks, particularly in specialized environments. To address this, this paper proposes a UAV-assisted mobile crowd sensing task assignment framework (UAM-TAF) based on reinforcement learning. First, in order to realize UAM-TAF, we introduce a Multi-Objective Task Assignment Goal (MOTGO) that innovatively accounts for factors such as cost, task scale, perception cost, and task diversity. Second, to achieve MOTGO with greater efficiency and precision, we propose a novel task assignment strategy based on reinforcement learning. In this strategy, the Priority and Expert Parametrized Deep Q-Network (PEP-DQN) is developed, building upon an enhanced P-DQN model for human-UAV collaboration. Having optimized the action parameter selection space in PEP-DQN to address the challenges of efficient continuous parameter selection and hybrid action space optimization. This approach effectively harnesses the complementary capabilities of human workers and UAVs to improve the completion rate and efficiency of perception tasks. Experimental results demonstrate that, compared to existing algorithms, the proposed framework has an improvement in perceived gain in the range of 15%-25% and converges faster and more consistently, based on considerations of equipment cost, task size, perception cost, and task variability. This provides a novel solution in the field of UAV-assisted mobile crowd sensing.
This paper focuses on cognitive radio (CR) for automated driving systems, which require stable and reliable wireless communications. Almost all relevant studies have adopted overly simplified models for vehicle-to-everything (V2X) communications. More realistic models for V2X are so topologically complex and dynamically variant that off-policy reinforcement learning (RL) algorithms cannot find stable solutions of optimization problems. To cope with the model problem, the CARLA simulator is combined with the SUMO simulator so as to simulate both microscopic and macroscopic traffic scenarios, in which parallel computing techniques can be used to speed up the training. Then, since on-policy RL algorithms such as proximal policy optimization (PPO) can find stable solutions to complex optimization problems, this paper expands PPO from a single-agent to a multi-agent version, which is referred to as SPMA-PPO, and combines SPMA-PPO with Transformers as its feature extractors, which is called TSPMA-PPO. Computer simulations of wireless resource scheduling with CR are conducted using the above testbeds, and demonstrate that TSPMA-PPO can achieve 45.1%, 10.9%, 7.8%, and 5.7% improvement in average lost data over the conventional off-policy RL algorithm, a proposed on-policy one with a CNN as its feature extractor, a previously proposed on-policy one without feature extractors, and a proposed on-policy one with an LSTM as its feature extractor, respectively.
Blockchain technology has emerged as a compelling architecture for electronic health record (EHR) management, offering decentralization, immutability, and cryptographic auditability to address longstanding vulnerabilities in centralized healthcare data systems. Yet these same architectural properties introduce privacy leakage risks that existing surveys have not systematically examined: transparency enables transaction pattern analysis and patient re-identification; immutability conflicts with regulatory erasure obligations; and consensus operations expose institutional identities. This systematic review, following PRISMA guidelines and analysing 47 peer-reviewed original articles published between 2023 and 2026, develops what is, to the authors’ knowledge, one of the first layered taxonomies of privacy leakage risks in blockchain-based healthcare systems, organized across the data, network, consensus, and application layers. The taxonomy maps fifteen distinct attack vectors to their architectural origins and available countermeasures. Findings reveal that advanced techniques, including attribute-based encryption with policy hiding, post-quantum lattice-based signatures, zero-knowledge proofs, and federated learning with differential privacy, provide significant but incomplete protection; critical vulnerabilities persist in metadata handling, peer-to-peer communication patterns, and consensus operations. Four concrete research priorities are identified: privacy-aware consensus protocols, standardized re-identification risk metrics, lightweight post-quantum primitives for resource-constrained IoT medical devices, and regulatory-compliant data erasure mechanisms compatible with GDPR and HIPAA.
With the rapid development of the Internet of Things (IoT), security issues in resource-constrained devices have become increasingly critical. Physical Unclonable Function (PUF) exploits unavoidable process variations introduced during IC manufacturing to generate unique responses and has been widely used for device authentication and cryptographic key generation. Among various PUF architectures, the Arbiter PUF (APUF) can generate a large number of challenge-response pairs (CRPs) with low hardware overhead. However, its reliability is highly sensitive to environmental variations and device aging, which may degrade the security of authentication protocols. In this paper, a challenge-response remapping reliability enhancement method for APUF based on adjacent challenge-bit flipping is proposed. First, a machine learning (ML) algorithm is employed to construct a software model of the APUF. A delay difference threshold is then introduced to identify challenges whose accumulated propagation delay differences fall below the threshold. Subsequently, specific adjacent bits in these challenges are flipped to transform them into highly reliable challenges. The proposed method is evaluated through Python-based simulations and FPGA-based hardware measurements. Experimental results show that the simulated reliability of the APUF can be improved to 100% under three environmental noise conditions. Moreover, within a temperature range from −20°C to 80°C, the measured reliability of the APUF can also reach 100% without reducing the number of available CRPs. Therefore, the proposed method significantly improves the reliability of APUFs without introducing additional hardware overhead or reducing the number of available CRPs, providing an effective solution for high-reliability authentication in IoT devices.
Magnetic indoor navigation remains challenging because indoor magnetic observations are affected by three coupled factors: information sparsity, perceptual aliasing, and sensor/behavioral heterogeneity. Deep Mag-Mamba introduces a magnetic-specific long-history folded-map token modeling pipeline for this setting: Temporal-Folding reorganizes the raw 4096-sample magnetic window into a 64 × 64 map without resampling, an EfficientFormer-style encoder with Decomposed Large-Kernel Attention (D-LKA) extracts compact folded-map tokens from the long history, and residual Mamba blocks model dependencies among these tokens for coordinate prediction. On the complete heterogeneous evaluation set, Deep Mag-Mamba achieves a mean absolute positioning error (MAE) of 0.342 m. On the Pocket-mode subset of the same heterogeneous evaluation, it obtains an MAE of 0.260m and a 95th-percentile circular error probable (CEP95) of 0.578 m. Under the strict Nexus 5X 11-anchor protocol, Deep Mag-Mamba achieves lower observed CEP95 than Res-T-LSTM in Calling, Handheld, and Pocket modes, demonstrating an empirical tail-error advantage in these modes, although Res-T-LSTM gives lower MAE. A trained-model efficiency benchmark further reports 67.72 frames per second (FPS) on an NVIDIA A800 GPU with a MIG 3g.40GB instance under the tested native input configuration.
Drone swarms are increasingly deployed for surveillance, targeting, and situational awareness, creating a growing need for effective countermeasures against adversarial aerial systems. This paper investigates the use of multimodal language models (MMLMs) with vision capabilities for leader drone identification in leader-follower drone swarms. Identifying leader drones enables effective disruption of adversarial swarm operations. A live drone swarm testbed was developed to collect video data across different leader positions, colors, and formations. The resulting dataset captures diverse swarm behaviors and enables evaluation under varying visual appearances and coordination patterns. Unlike approaches that rely on telemetry or communication signals, which may be unavailable or encrypted in adversarial environments, the proposed method operates solely on externally observable video data. We evaluate (a) whether off-the-shelf MMLMs can identify leader drones, (b) the impact of visual-temporal input, and (c) the effect of parameter-efficient fine-tuning under computational constraints. Results show that zero-shot MMLMs perform near random-guessing accuracy, whereas fine-tuned models achieve substantial improvements. We compare the proposed framework against two baselines. The first is a telemetry-derived trajectory heuristic that achieves 49.02% accuracy when structured trajectory information is available, although such telemetry may be inaccessible in adversarial/military environments. The second is a detector-tracker-heuristic pipeline based on YOLO, ByteTrack, and trajectory-based leader-identification heuristics, achieving 31.91% accuracy. In comparison, the fine-tuned Qwen2.5-VL model achieves 45% accuracy using only video observations. These findings demonstrate the potential of MMLMs for practical leader identification in drone swarms without requiring access to telemetry or communication signals.
Edge-assisted Internet of Things (IoT) systems require collaborative authorization that balances threshold-based coordination with fine-grained policy enforcement. However, existing threshold signatures lack expressive policies and accountability, while attribute-based signatures suffer from inefficient multi-party aggregation. To bridge this gap, we propose PA-ATAS, a policy-aware accountable threshold aggregate signature framework. PA-ATAS seamlessly integrates threshold BLS signatures, attribute-based signatures (ABS), and simulation-extractable zero-knowledge proofs. In this framework, IoT devices generate lightweight local authorization transcripts, while an untrusted edge gateway compresses these contributions into a single recursive global policy-binding proof. This decoupled architecture achieves strictly O(1) public verification complexity and constant signature size, utilizing a hash anchor to maintain the linearly growing accountability payload off-chain. We formally prove that PA-ATAS satisfies unforgeability, policy soundness, privacy, and traceability against malicious gateways and adaptively corrupted devices. Comprehensive experiments demonstrate that offloading the aggregation burden to the edge ensures highly efficient and scalable public verification, making PA-ATAS highly practical for resource-constrained IoT deployments.
Federated Learning (FL) has emerged as an important paradigm for distributed intelligence in large-scale AI-based IoT systems under the Integrated Sensing, Memory, Communication, and Computation (SMCC) framework. However, its privacy guarantees may be threatened by gradient leakage attacks under dynamic training conditions. Existing client-side attacks often suffer from limited reconstruction fidelity and high computational overhead. In this paper, we propose PAFS, a generative gradient leakage framework based on Poisoning-driven Analytical Feature Separation, designed to investigate privacy vulnerabilities relevant to SMCC-enabled collaborative learning. PAFS introduces a feature separation mechanism in which strategically poisoned updates amplify target-class gradients, enabling target-related features to be extracted from fully connected (FC) layers. To bridge the semantic–spatial gap in reconstruction, we further introduce a Hierarchical Feature Fidelity (HFF) mechanism that enforces multi-scale feature consistency to preserve both fine-grained textures and semantic structure. As a result, target images can be reconstructed via a single forward pass. Experiments on CIFAR100 and ImageNet show that PAFS achieves competitive reconstruction quality with improved efficiency compared with representative baselines, while maintaining stable performance under varying client scales and Byzantine-robust aggregation rules, and remaining effective under low-magnitude gradient perturbations. The results highlight potential privacy risks in SMCC-enabled FL systems and provide insights for privacy evaluation in intelligent IoT systems.
The YOLO series has achieved remarkable progress in real-time object detection, yet nano-scale variants remain constrained by single-round feature pyramid topologies where the global correlation prior is distributed only once and extreme pyramid levels exchange information indirectly. In this paper, we propose YOLO-Prism, an efficient nano-scale detector built upon YOLOv13 that addresses this single-round limitation through asymmetric multi-round refinement. We propose a Multi-Round Feature Pyramid Network (MR-FPN) that exploits the decoupling between YOLOv13’s precomputed correlation prior and its lightweight gated distribution mechanism, appending a secondary top-down refinement pass at only ∼6% parameter overhead under a “compute once, distribute many” paradigm. Subsequently, as a near-zero-cost complement, we selectively integrate DCT-based Multi-Spectral Channel Attention (MSCA) at gradient-identified critical layers, contributing an additional gain at merely 0.01M parameters. Extensive experiments on MS COCO, PAS-CAL VOC, and VisDrone benchmarks demonstrate that YOLO-xsPrism achieves 41.66% AP50:95 on COCO val2017, surpassing the strong YOLOv13n baseline by 0.62 points under identical training conditions and outperforming YOLOv12n by 1.10 points, at 2.63M parameters and 7.1 GFLOPs. On YOLOv11n, whose neck lacks such a prior, the same asymmetric topology yields 0.50 AP50:95 at 3.8% overhead, confirming that the benefit of multi-round refinement is independent of any specific gating mechanism. This lightweight algorithm can be applied to intelligent video measurement, industrial edge detection sensors, and IoT monitoring.
Dynamic metasurface antennas (DMAs) are emerging as a promising technology for future satellite communications, offering reductions in power consumption and hardware costs. However, the proliferation of Internet of Things (IoT) applications has exacerbated spectrum scarcity in DMA-assisted satellite communication systems. To tackle this limitation, we propose a novel integrated reconfigurable intelligent surface (RIS) and DMA empowered satellite symbiotic radio (SR). Specifically, the satellite equipped with a DMA transmits signals to primary users with the assistance of the RIS. Meanwhile, following the RIS-assisted SR principle, the RIS transmits its own signal to the secondary user. Considering low-resolution digital-to-analog converters (DACs) at the transmitter, we investigate the maximization of the weighted sum rate (WSR) through the joint optimization of the transmit beamforming vectors, the weight matrix associated with the DMA, and the phase shifts of the RIS. The strong coupling of optimization variables and the structural limitations inherent to the DMA make the problem highly challenging. To overcome these difficulties, we develop an efficient alternating optimization method based on the Lagrangian dual transform and quadratic transform algorithms to reformulate the original problem as a computationally tractable form. Then, we apply the penalty convex-concave procedure principle and the complex circle manifold algorithm to design the RIS phase shifts and the DMA weight matrix, respectively. In addition, an energy efficiency (EE) maximization formulation is developed to further evaluate the power-consumption benefit of the DMA-assisted transmitter. Numerical results demonstrate the effectiveness of the developed algorithm and highlight the favorable rate-power tradeoff of the DMA-assisted architecture, which achieves competitive WSR with improved EE compared with the conventional full-digital scheme.
