The Internet of Vehicles (IoV) enhances road safety through cooperative communications and the broadcasting of Basic Safety Messages (BSMs). However, the dynamic network of IoV architectures makes them vulnerable to misbehaviors such as position malfunctions. Existing detection methods are trained on labeled data, resulting in low accuracy in detecting complex and unknown attacks. This work proposes the Unsupervised Transformer-based Misbehavior Detection (UTMD) method. UTMD employs a dual-branch architecture that independently encodes physical motion and communication behavior, these representations are then integrated using a cross-attention fusion mechanism. Misbehavior is identified by a dynamic-threshold decision module that uses an MLP to adaptively distinguish genuine behavior from misbehavior based on statistical distributions. The proposed model is evaluated using the VeReMi_V2 dataset, training uses only genuine data, while testing uses mixed data. The results demonstrate the exceptional effectiveness of UTMD, achieving Acc of 99.40

The Internet of Vehicles (IoV) enables efficient traffic information sharing but is highly vulnerable to malicious attacks, false messages, and unreliable data dissemination. Existing trust models often suffer from static evaluation limitations, high communication overhead, and consensus bottlenecks. To secure the IoV system, this paper proposes a dynamic reputation framework based on deep learning and a hybrid blockchain architecture. Specifically, we design a data-driven closed-loop trust mechanism. At the edge layer, Roadside Units (RSUs) employ a pre-trained Gated Recurrent Unit-Fully Connected Network (GRU-FCN) to evaluate message credibility and quantify prediction uncertainty. By extracting spatiotemporal features from historical report sequences, this model effectively captures the dynamic temporal coherence of vehicle behaviors. This is combined with a Bayesian inference model that fuses multi-source reports using historical reputations as weights to determine event authenticity. Subsequently, a GRU-FCN and Bayesian-based Reputation Consensus (GBRC) mechanism is executed within a hybrid blockchain, which anchors local private chain records to a global public chain. Smart contracts are then automatically triggered to penalize or reward vehicle reputations based on the consensus outcome, forming a continuous dynamic feedback loop. Extensive experiments demonstrate that the proposed framework maintains high precision and recall against various malicious attacks. Furthermore, it significantly reduces communication overhead to linear growth, ensuring high throughput and low latency in dense vehicular networks.

Existing novel view synthesis methods based on Neural Radiance Field (NeRF), a prominent artificial intelligence approach, often suffer from prolonged training. While Three-dimensional Gaussian Splatting (3DGS) is faster, both methods struggle in sparse-view scenarios, producing geometric inconsistencies. In 3DGS, this stems from sparse initial three-dimensional points and flawed gradient-based densification, yielding poorly positioned or shaped Gaussians. To address these limitations, we propose a framework for pseudo-view-driven Gaussian optimization. Our cohesive strategy first employs pseudo-view Gaussian densification. This core module leverages a Multi-View Stereo (MVS) network, empowered by adaptively generated pseudo-views, to overcome sparse initialization and create a dense, geometrically accurate set of Gaussians. The process is then refined by the abnormal Gaussian shape correction module, a regularization that analyzes multi-view consistency across these pseudo-views to identify and correct misshapen primitives, ensuring high rendering fidelity. Experimental evaluations on public datasets show our method outperforms both state-of-the-art 3DGS and NeRF approaches, achieving competitive rendering quality and geometric fidelity.

Low-frequency periodic trends prevalent in traffic flow patterns tend to obscure multi-scale variations, while random high-frequency fluctuations triggered by sudden events further exacerbate modeling instability. Existing attention mechanisms and graph-based models achieve partial improvements but still struggle to capture multi-scale temporal patterns and dynamic spatial correlations. To address these challenges, we propose WaveGFormer, a wavelet-enhanced spatio-temporal graph transformer for traffic flow prediction. The model employs discrete wavelet decomposition (DWT) to separate trend and disturbance components, enabling adaptive multi-scale temporal modeling through learnable filters and multi-head attention. Then, we designed a dual-channel architecture for the spatio-temporal encoder. The temporal channel utilizes wavelet-enhanced self-attention for learning long-term dependencies, while the spatial channel integrates graph convolution and spatial self-attention to jointly capture local and global correlations based on static and dynamic adjacency matrices. Finally, a gated fusion module adaptively combines the outputs from both channels to form robust spatio-temporal embeddings. Evaluations on four real-world PeMS benchmarks show that WaveGFormer consistently surpasses competing methods, achieving up to 4.74% lower MAE on PeMS08, verifying its effectiveness and superiority for robust traffic flow forecasting.

Personalized Federated Learning (PFL) based on model decoupling has emerged as a prominent paradigm for mitigating statistical heterogeneity. However, most existing methods rely on static or over-simplified local representations, neglecting the severe feature misalignment during global aggregation. This inevitably precipitates representation degradation and convergence difficulties, especially under highly heterogeneous data distributions. To this end, we propose a novel framework termed PFL-AP for adaptive client representation calibration and knowledge fusion. Specifically, we design an Adaptive Affine Calibration (AAC) mechanism that empowers each client with affine parameters to jointly calibrate the scale and shift of shared representations, thereby explicitly modeling client-specific distribution shifts and mitigating the performance degradation caused by global-local representation mismatch. In addition, we introduce a Prototype-driven Knowledge Fusion (PKF) mechanism that reformulates client-specific shifting factors as semantic prototypes to perform cross-client knowledge filtering based on representation similarity. By selectively integrating only semantically consistent and beneficial knowledge, PKF facilitates constructive collaboration while suppressing interference. Extensive experiments on five public benchmarks demonstrate that PFL-AP achieves superior representation alignment and exhibits strong generalizability and robustness in tackling data heterogeneity.

Multi-agent trajectory prediction relies heavily on information exchange, where incomplete state sharing, limited interaction frequency, and communication uncertainty readily induce collaborative instability. Current approaches primarily rely on heuristic interaction rules, failing to explicitly quantify the trade-off between communication overhead and prediction gains. This implicit assumption often leads to suboptimal information exchange and lacks evolutionary stability for communication strategies in dynamic traffic environments. Specifically, we propose the Stability-aware Communication via Weighted Potential Games (SC-WPG) framework. First, we leverage the continuous-time dynamics of Liquid Neural Networks (LNNs) to design a dynamic phase-coupling mechanism that aligns the temporal evolution of agent states, alleviating strategy asynchrony. Second, we establish a liquid state-sharing pool to exchange compact latent representations encoded from local observations, while promoting efficient implicit state fusion through spatio-temporal attention mechanisms. Finally, communication scheduling is formulated as a weighted potential game to explicitly quantify the trade-off between gains and costs. A stability-aware weight adaptation mechanism based on latent-state variation is introduced to regularize communication behavior under dynamic constraints. Experimental results on the NGSIM and HighD datasets show that SC-WPG achieves the lowest average root mean square error (RMSE) of 0.89 and 0.55, respectively. These results correspond to improvements of 10.1% and 6.8% over the strongest competing baselines under the same communication-constrained setting.

Improving energy efficiency is critical for the sustainable development of urban public transportation. Regenerative braking is widely employed in urban rail transit to recycle braking kinetic energy into the traction network, thereby enhancing system efficiency. However, without effective scheduling, excessive feedback energy can induce instantaneous voltage spikes, leading to line overheating and accelerated equipment aging. Existing studies often fail to fully address these challenges due to simplified physical models and limited adaptability to real-time environments. To overcome these limitations, this study proposes a dynamic scheduling method for the efficient utilization of regenerative energy within a train fleet. A physical simulation system featuring a "Network-Train-Control" three-layer architecture is constructed. By formally describing the physical coupling among network topology, operational rules, and train kinematics, the system enables accurate energy profiling under realistic impedance and signaling constraints. Furthermore, a finite state automaton (FSA) is utilized to abstract continuous train dynamics into discrete states, facilitating a braking-event-triggered Model Predictive Control (MPC) framework. This framework predicts and dynamically adjusts fleet operations within a receding horizon to maximize the immediate absorption of regenerative energy. Experimental results demonstrate that the proposed method achieves active energy cooperation among trains, increasing the regenerative energy utilization rate by approximately 11%, thereby offering a viable technical solution for low-carbon urban transit.

Federated learning provides a new paradigm for secure data exchange and collaborative computation. However, it still faces critical challenges such as data heterogeneity and client drift. While integrating hypernetworks or Transformer architectures into federated learning addresses data heterogeneity and enhances model generalization, issues such as low communication efficiency and inadequate capture of fine-grained details remain unresolved. To address these limitations, we propose a personalized federated learning method based on deep multisampling and hypernetwork dynamic adaptation (Per-FedDMA), which aims to balance the universality and personalized ability of federated learning in heterogeneous scenarios. First of all, a hypernetwork is introduced to generate the weight parameters of the Transformer architecture to dynamically adapt to the data distribution of different clients. Secondly, the deep multisampling algorithm (DMA) is designed to fuse local and global information through multi-path and multi-scale feature extraction mechanisms, so as to extract rich image features under different sensory fields and enhance the ability to capture low-level details of images. In addition, Per-FedDMA employs an adaptive optimization strategy to accelerate model convergence while improving communication efficiency. This approach enhances the robustness and generalization capability of federated learning in heterogeneous environments. Experimental results show that Per-FedDMA improves 1.63% and 8.80% compared to the baseline methods on the two benchmark datasets and achieves a breakthrough in terms of communication efficiency.

Accurate trajectory prediction for autonomous driving and intelligent transportation systems remains challenging due to dynamic traffic environments, complex multi-agent interactions and privacy-constrained decentralized data settings. The key difficulty lies in capturing uncertain agent behaviors while maintaining robust model adaptation under heterogeneous traffic data and data locality constraints. To address this, we propose a novel multi-agent collaborative inverse reinforcement incremental federated learning (MAC-IRIFL) framework. Firstly, an extensible reward alignment mechanism is developed. It improves cross-agent reward consistency under heterogeneous multi-agent driving data through a Kullback-Leibler (KL) divergence metric based on a quadratic Euclidean surrogate and a low-rank singular value decomposition (SVD) pivot. Secondly, an incremental aggregation scheme with parameter freezing is designed for dynamically participating clients. It supports privacy preserving collaborative training by retaining raw trajectory data locally at each agent. Meanwhile, local feature extractors remain specialized and only high-level decision layers are synchronized. Finally, a gated recurrent unit (GRU) based encoder-decoder conditioned on semantic maps is constructed. It better captures temporal dependencies, scene context, and future trajectory evolution. Experiments on the Lyft Level 5 (Lyft5) and nuScenes datasets demonstrate that MAC-IRIFL achieves superior robustness and prediction accuracy compared with state-of-the-art methods.

Existing blockchain-based covert communication schemes encounter persistent challenges, including inadequate security guarantees and limited embedding capacity. Additionally, the parties involved in such communications often rely on static keys for extended periods, and multiple transaction addresses demonstrate inherent correlations due to their derivation relationships or fund flows. This vulnerability enables third-party attackers to trace all related addresses once a specific covert transaction is exposed. To address these issues, we introduce a covert communication model based on multi-channel and vector space coding, featuring three key innovations. Firstly, we organize communication addresses into multiple groups and develop an address space coding strategy based on vector space coding. Secondly, we integrate transaction amounts and transaction IDs with transaction address to facilitate the transmission of single-session communication keys. Finally, we embed pseudorandom ciphertext into digital signature channel and achieve a high embedding capacity per transaction by employing ring signatures and UTXO structures. Moreover, we provide a theoretical analysis evaluating various aspects of the model, including tamper resistance, detection resistance, correlation resistance, and scalability. Performance validation is further supported by experiments assessing efficiency, embedding rate, ciphertext randomness, and the impact of the number of groups.

The openness and traceability of blockchain systems pose inherent challenges to covert communication, as communication identities are easily linked and embedded patterns are susceptible to detection. Most existing blockchain-based covert communication schemes mainly focus on on-chain data embedding, while paying limited attention to the anonymity of communication paths.To address this issue, this paper proposes a blockchain covert communication model that integrates onion routing with dynamic transaction embedding. The Tor network is used to anonymize transaction broadcasting paths, effectively decoupling off-chain communication sources from onchain transaction activities. In addition, a dynamic embedding position mapping and amount-interval encoding scheme based on HMAC-SHA256 is adopted to disperse covert information into ordinary transactions, improving resistance to statistical analysis while preserving transaction semantics. Experimental evaluation is conducted on a P2P topology that resembles a consortium blockchain environment. The results show that, compared with direct communication and random forwarding, the proposed scheme achieves higher anonymity and better embedding stealthiness, while maintaining acceptable communication efficiency and deployability.

Federated Learning (FL) enables the collaborative training of a global model across multiple clients, effectively preserving data privacy and overcoming data silos. However, in real-world scenarios, data is often Non-Independent and Identically Distributed (Non-IID), which can lead to model drift and unstable convergence. Traditional client selection methods struggle to adapt to dynamic training environments, inherently limiting model performance improvements. Simply introducing new clients into training can also disrupt the trajectory of the global model. To address these challenges, we propose FedAEDC, an adaptive client selection method based on distillation calibration for federated learning. FedAEDC evaluates the value of client data by considering alignment, which measures the directional consistency of updates, and contribution, which assesses the informational richness of model updates. Furthermore, the method dynamically adjusts the number of participating clients by monitoring the rate of change in loss and the variance in aggregated gradients during training. Additionally, it employs a value-based teacher-student framework to distill and calibrate the updates from newly added clients. This mechanism mitigates model drift while preserving data diversity, leading to improvements in both model accuracy and training efficiency. Experimental results show that FedAEDC enhances classification accuracy and reduces training time overhead, achieving a better balance between model performance and communication efficiency.

Retrieval-augmented generation (RAG) improves factual grounding by conditioning generation on external evidence, but it also makes answer quality more dependent on retrieval robustness. Under poisoned retrieval, adversarial passages may survive ranking and enter the final prompt, creating a reliability risk even when the generator itself is unchanged. Graph-based reranking is a promising retrieval-stage defense because it can exploit structural relations among retrieved candidates. However, our analysis shows that this defense line remains vulnerable when poisoned passages form small near-duplicate groups that are both strongly query-supported and locally cohesive. To address this residual failure mode, we propose GHP (graph hardening penalty). Built on a locally restored HRSIM graph-based reranking baseline, GHP operates at a fixed reranking-stage interface and leaves retrieval and generation unchanged, while demoting candidates that jointly exhibit strong query support and strong local cohesion. We also introduce GHP-Lite, a selective inspection-scope variant that narrows the suspiciousness-inspection scope while preserving the same fixed final top-5 evidence interface. Across the main attacked-setting matrix, spanning two backbones, three datasets, and two poisoning attacks under one shared local evaluation family, the restored HRSIM baseline consistently improves over no defense and serves as a credible graph-based reference line. On the NQ anchor line with Qwen2.5-7B-Instruct under LM-targeted poisoning, HRSIM reduces attack success rate (ASR) from 50.3% to 5.3% and improves exact match (EM) from 26.7% to 41.3%. GHP-R further reduces ASR to 1.0% while increasing EM to 44.7%, and a pooled query-level significance check shows that its robustness gain over restored HRSIM is statistically significant on this anchor setting. GHP-Lite closely preserves the milder comparison point under a narrower inspection scope, while a simple TF-IDF cosine near-duplicate suppression baseline remains close to the no-defense line and therefore does not account for the gains of the graph-based line. Benign full-corpus evaluation shows that clean-side behavior varies substantially across datasets rather than following a uniform non-degradation pattern, with the largest penalty concentrated on HotpotQA. Supplementary transfer and attack-diversity checks broaden the empirical picture, but remain supportive rather than conclusive. Overall, the results suggest that GHP is a practical failure-mode-driven hardening extension for restored graph-based poisoned-RAG defense under the current local setting, while also highlighting the importance of operating-point selection across different evaluation conditions.

The Whale Optimization Algorithm (WOA) is a meta-heuristic algorithm widely utilized in the field of engineering optimization. However, it suffers from defects such as low solution accuracy, slow convergence, and a tendency to fall into local optima. To overcome these shortcomings, this paper proposes a novel approach named EGRMWOA. First, we design an adaptive elite weight mechanism to achieve a balance between global exploration and local exploitation by dynamically adjusting the influence of the elite solution. Second, by introducing a local grey wolf optimal regulation mechanism, a refined search is conducted around the current optimal solution, enhancing solution accuracy and local exploitation capability. Moreover, we improve the random search formula to accelerate convergence. Finally, we introduce a similarity elimination and perturbation mutation strategy, which increases population diversity and enhances the ability to escape local optima. Experimental results on 23 standard test functions and the CEC2019 benchmark demonstrate the significant advantages of EGRMWOA. Compared to five frontier WOA variants, its average ranking reached 1.63 and 1.35, respectively. Furthermore, it maintained excellent average rankings of 1.804 and 2.1 against six other well-known meta-heuristic algorithms, securing the first-place rank overall. In 15 real-world engineering optimization problems, EGRMWOA demonstrates exceptional solving capabilities with an average ranking of 2.833. Crucially, in the practical optimization problem of electric vehicle charging and swapping scheduling, its performance notably surpasses that of the standard WOA and GWO.

Transportation mode detection (TMD) is essential for travel behavior analysis and intelligent transportation systems. Existing deep learning methods suffer from a speed-accuracy trade-off: Transformer-based models incur quadratic time complexity with respect to sequence length, leading to excessively long training and prediction time, while MLP classifiers with fixed activation functions lack expressiveness for discriminating modes with similar characteristics. We propose TMD-BMKAN to address both issues. The model employs a Bidirectional Mamba encoder with linear time complexity for global context modeling to reduce training and prediction time, and a Kolmogorov-Arnold Network (KAN) decoder with learnable B-spline activations for adaptive decision boundary learning to improve accuracy on similar modes. Experiments on SHL and HTC show that TMD-BMKAN reaches 98.60

Origin-Destination (OD) demand prediction is critical in intelligent transportation systems, helping alleviate congestion and optimize traffic. However, current methods respond poorly to complex patterns and unexpected events, often causing inaccurate predictions. Their reliance on fixed-interval time slots also limits the capture of multi-scale traffic dynamics. To address these issues, we propose an environment tree-guided semantic random walk, introducing semantic similarity constraints into node transitions. This lets node embeddings capture both topological and semantic information for improved cross-regional traffic modeling. Our hierarchical memory system integrates continuous, discrete, and reflective memory for unified, multi-scale representations of short-term dynamics, long-term patterns, and semantic context. Based on this, we present an OD prediction method, ETSRW-HM (EnvTree-Guided Semantic Random Walk and Hierarchical Memory). Experiments show our model surpasses mainstream baselines in RMSE and PCC metrics (1.5327 and 0.8611), demonstrating greater accuracy and trend prediction.

As urban rail transit systems evolve within the Industrial Internet of Things (IIoT), the intelligent recovery of regenerative braking energy becomes critical for energy efficiency. However, the existing train operation optimizations primarily focus on time-domain synchronization, frequently neglecting the spatial impedance constraints of the DC traction network. This oversight creates a discrepancy between theoretical energy matching and actual absorption. To address this, this paper proposes a spatiotemporal synergistic optimization framework integrating the analysis of electrical energy transmission factors and train relative motion. First, a dynamic multi-node circuit model based on Kirchhoff’s laws is established to characterize train fleet operations. By evaluating electrical energy transmission factors, the current distribution ratio and line impedance loss are identified as primary determinants of absorption efficiency. This physically quantifies the coupling among instantaneous energy distribution, transmission loss, and source-load relative distance. Second, a time-domain integration-based gradient analysis framework is formulated to deconstruct the energy gradient into amplitude and directional components. By mapping the relative position and speed of interacting trains, their relative motion states are systematically categorized. Subsequently, an adaptive gradient optimization strategy based on these motion states is introduced, which fine-tunes dwell times to precisely guide train trajectories into a low-impedance “optimal window” for energy absorption. Finally, a case study using operational data from Luoyang Metro Line 1 validates the proposed framework. Results demonstrate that the framework achieves dual spatiotemporal matching of braking and traction trains, outperforming the traditional fixed timetable and improving the regenerative braking energy absorption rate by approximately 13%.

Vehicle trajectory prediction plays a crucial role in IoT-based intelligent transportation systems, which can effectively address key issues, such as driving safety and multivehicle collaboration. However, the sensitivity of trajectory data and the reluctance of data holders to share it constrain the prediction model's ability to capture vehicle behavior patterns in different scenarios. To address the above problems, we propose a blockchain-enabled asynchronous federated proximal policy optimization framework (BE-AFPPO) for the trajectory prediction of self-driving vehicles. First, we propose a curiosity proximal policy optimization (C-PPO) algorithm. The method utilizes a driven exploration strategy to actively motivate the intelligent agent to explore the unknown state space. The avoidance policy model reaches a local optimum when processing trajectory data. In addition, we design historical gated recurrent unit (GRU) and future GRU as input layers. The target's historical motion features and future trajectory features are extracted, respectively. Then, various data is received through asynchronous federated learning. This model can fully learn the vehicle's behavior patterns in different scenarios, which improves prediction accuracy. Based on this, we develop a blockchain-based dynamic group practical Byzantine fault tolerance (DG-PBFT) consensus algorithm. This enhances the credibility and integrity of the data while enriching the sources of trajectory data. Finally, we perform the experiments on the publicly available dataset nuScenes. The results demonstrate that the proposed method improves the robustness and accuracy of trajectory prediction.

Accurate traffic flow prediction is crucial for alleviating traffic congestion and optimizing intelligent transportation system (ITS). However, traffic flow is subject to uncertainties and exhibits complex spatial and temporal dependence and dynamic change characteristics. Moreover, many efforts rely on a single view, which makes it difficult to comprehensively capture multiple levels of spatial and temporal correlations, thus limiting the accuracy of predictions. Therefore, we propose the multiview spatiotemporal dynamic graph convolution framework MVSTDG for more comprehensively exploring and fusing the multiview spatiotemporal features. First, we design a dual-path time-patch convolution module (TPConv) module to separately model short-term fluctuations and long-term periodic trends, enabling effective extraction of dynamic features at multiple temporal scales. Second, we construct a data-driven traffic pattern library to generate dynamic adjacency matrices and integrate them with static topologies view. An adaptive diffusion graph convolutional network (ADGCN) is then employed to model both local and global spatial correlations. In addition, we design a cross-gated spatiotemporal fusion mechanism that adaptively adjusts the contribution of short-term and long-term information, enhances the interaction of spatiotemporal information, and improves the model's adaptive capability under different time scales. The experimental results show that MVSTDG outperforms the state-of-the-art baselines in several evaluation metrics and demonstrates higher prediction accuracy and stability on the four real datasets.

Federated learning is a promising privacy-preserving paradigm for collaboratively training global machine learning models across distributed devices. However, traditional federated learning encounters difficulties in training global model and issues of single-point failure when faced with complex and diverse distributed data. In this paper, we propose a decentralized federated learning approach based on meta-learning and equalized multi-scale feature fusion to enhance the accuracy of the classification task. Firstly, we present a platform architecture of blockchain-based federated learning systems for data sharing among clients, which achieves a trusted decentralized framework. In this architecture, we construct personalized federated learning models tailored to individual users by utilizing meta-learning techniques. Secondly, to address the challenge of fusing complex data features from local users, we formulate a novel equalized multi-scale feature fusion method that employs the dual attention mechanism to extract local data features in a fine-grained manner. Experimental results demonstrate that the proposed approach improves the accuracy by approximately 2% and reduces the loss by approximately 0.25 compared to mainstream federated learning methods.