Image data trading in the Internet of Vehicles (IoV) can improve data sharing and service quality, but it also faces copyright protection and transaction fairness challenges because image data can be easily copied, modified, and illegally resold. Existing watermarking- and fingerprinting-based copyright protection methods may distort traded images, while conventional perceptual hashing methods may misclassify independently captured near-duplicate images as pirated copies. To address these issues, this paper proposes a blockchain-based image data trading scheme that supports copyright-preserving and fair trading in IoV. The proposed scheme stores perceptual hash sequences on the blockchain as compact copyright information and uses smart contracts to verify copyright, distribute rewards, and punish malicious behavior during the transaction process. We further design a perceptual hashing algorithm that identifies copy-attack images while distinguishing them from near-duplicate similar images. Experimental results show that the proposed algorithm achieves a copied-vs-near-duplicate AUC of 0.9925 and an equal error rate of 2.91%. The ablation study further confirms the effectiveness of the near-duplicate separation term, whose removal increases the near-duplicate false positive rate from 3.00% to 18.68%. Smart-contract evaluation shows that the proposed scheme can support copyright-aware fair trading with acceptable gas consumption.
The field of recommender systems is currently seeing a lot of attention around the topic of multi-behavior recommendation. This approach is good at solving data sparsity and cold start issues compared to single-behavior recommendation. However, the existing methods are still facing the following key challenges in representing multi-behavior semantic information: (i) While current recommender systems predominantly rely on user-item interactions for modeling, they may overlook potential semantic knowledge, thereby limiting their effectiveness in sparse data scenarios. (ii) Compressing multiple behaviors into a single embedding space and neglecting semantically similar users in contrastive learning process jointly distort interactive semantic representations, ultimately impairing the modeling of fine-grained user preferences and breaking the continuity of semantic structure in embedding space. To cope with the above challenges, we design Multi-view Semantic Knowledge enhancement for Multi-behavior Recommendation (MSKMR) framework to capture diverse semantic signals through three semantic knowledge views, i.e., the multi-behavior user-item interaction view, the item-item relation knowledge view and the user-user relation knowledge view. The first view introduces behavior-supervised graph convolutional encoder and semantic-aware contrastive learning to fully mine the user preference semantics under each behavior. The latter two views leverage a knowledge graph generator and a behavior-projected knowledge encoder to capture more potential semantic knowledge. Finally, contrastive learning is employed to align and integrate embeddings from the three semantic knowledge views. Across four real-world datasets, MSKMR has been proven to surpass a variety of state-of-the-art recommendation methods.
This paper presents an effective hierarchical decoder supervision framework for 3D medical image segmentation that introduces a Hierarchical Decoder Supervision mechanism to improve the SwinUNETR Pre-training model. While the SwinUNETR leverages the Swin Transformer as its encoder, it can suffer from limited gradient flow during pre-training due to its simplistic linear decoder. The proposed method overcomes this by constructing multi-level decoders that integrate features from different hierarchical levels via skip connections and reconstruct images at multiple scales. A hierarchical weighted reconstruction loss is also proposed to guide the learning process across all decoder layers, enabling effective feature alignment and better representation learning. Experiments conducted on multiple datasets, including the Beyond the Cranial Vault (BTCV) datasets, Automated Cardiac Diagnosis Challenge (ACDC) datasets and the Brain Tumor Segmentation (BraTs), demonstrate that this approach consistently improves segmentation performance over strong transformer-based baselines across multiple datasets, especially in preserving edge continuity and segmenting small or complex anatomical structures. This work highlights the benefits of deep supervision in self-supervised pre-training for volumetric medical image analysis.
Safety stock (SS), encompassing reorder point (RP) and maximum stock (MS), is designed to minimize inventory cost while maintaining maintenance safety for manufacturing enterprises. In recent years, advanced computational techniques have been applied to forecast spare parts demand (SPD), enabling dynamic updates to SS settings. However, existing methods have often failed to produce reliable SS decisions when confronted with significant demand volatility and random failures. The reliability of SS decisions remains an open problem and is still in its early stages of investigation. To address this issue, in this paper, a reliable SS model with uncertainty evaluation of SPD is proposed in this paper. Initially, a basic setting of RP and MS is derived on a three-level warehousing architecture by simultaneously minimizing excess inventory cost and shortage cost. Building upon this foundation, a novel SS reliability metric is introduced by integrating static information (inventory coverage and emergency replenishment rate) with dynamic information (decision compactness ratio). To enhance the decision compactness ratio, a multivariate intermittent time series forecasting method based on an improved graph neural network is developed. Accurate forecasting with a confidence interval is achieved through bootstrap resampling. Finally, a reliable RP and MS configuration is obtained by incorporating the forecasting interval into the basic stock setting. Validation is conducted using an actual spare parts dataset from a large rail transit manufacturing enterprise in China. The experimental results demonstrate that the proposed model not only achieves higher SPD prediction accuracy but also improves inventory turnover and coverage while significantly enhancing reliability.
With the surge of high-dimensional data, feature selection has become an indispensable and effective method to distill informative and interpretable features by leveraging the underlying graph structure. However, the graph learning has been suffered from the irrelevant and redundant information on real-world data, while the nonlinear data structure and the ambiguous membership in subspace has not been well delved yet, which result in suboptimal feature subset.To address these drawbacks, we present a sparse and fuzzy subspace learning method via fuzzy K-means and anchor graph (FKAG) embedding with $l_{2,0}$-norm row-sparsity for unsupervised feature selection. In particular, fuzzy K-means clustering and anchor graph learning is jointly considered to efficiently preserve manifold information into the fuzzy label space. The proposed method can simultaneously obtain fuzzy membership for data samples and optimal subspace distribution with consensus row-sparse projection. To solve the challenging nonconvex objective function with semi-orthogonal and $l_{2,0}$-norm constraints, an alternative optimization method is devised to solve the challenging problem with comprehensive theoretical analysis. Extensive experiments on three synthetic data and eight real-world datasets illustrate the effectiveness and superiority of our approach in comparison with several state-of-the-art methods.
Deep reinforcement learning (DRL) has achieved remarkable breakthroughs and even matched human intelligence in the tasks related to discrete-time systems such as Atari games. However, DRL in continuous-time systems faces significant challenges because real-world scenarios often involve dynamic systems governed by differential equations, which have stringent safety requirements. To capture the dynamic properties of these systems accurately, the method of reducing the sampling interval is often used. Unfortunately, the reduction of sampling interval causes high-freuency sampling, which leads to a large volume of redundant experience samples that continuous-time reinforcement learning must cope with under a fixed and often unavoidable sampling interval. To solve this problem, in this paper, we introduce a Continuous-time Priority experience replay Hamilton-Jacobi-Bellman deep Q-learning Network (CPHQN), which aims to maximize sample efficiency and learning effectiveness in continuous-time systems with fixed sampling intervals by learning to prioritize redundant transitions through priority optimization, thereby emphasizing those uncommon but very valuable experience data. Meanwhile, it significantly enhances the learning process for experiences with large temporal difference errors amidst redundant data. Furthermore, semi-Markov decision process modeling is employed to achieve deep Q network (DQN) fitting in continuous-time systems. We conducted extensive experiments across six tasks in Mujoco, and the experimental results have demonstrated that CPHQN effectively addresses the problem of temporal difference (TD) errors becoming insufficiently informative when learning from highly redundant transitions generated under a fixed sampling interval, and this novel method surpasses baseline methods in various tasks while maintaining low action variance.
This work considers black-box approaches to categorical optimization where the decision variables take values within a finite unordered set. We propose an efficient probability modeling method to parameterize the problem and perform the search using the evolution strategy framework. The new method involves information-geometric techniques to explore promising regions while being able to control step-sizes automatically. We show that the new method converges on a smooth surrogate of the original categorical problem and can approach a neighborhood of its first-order optimal solutions. Several benchmark problems and a real-world task in natural language processing, namely the discrete prompt search, are chosen for empirical study. The results demonstrate superior performance of the new method in terms of speed and solution quality compared with existing algorithms.
Graph Neural Networks (GNNs) have demonstrated their effectiveness in various real-world applications on complex non-Euclidean data, such as semi-supervised classification of social networks, molecules, and recommender systems. Substantial paradigms have been proposed to exploit the potentialities contained in graph topologies. However, the essential, high-level abstracted, but universal pattern of graphs, especially how information flows among nodes and edges, is still unrecognized. To remedy the deficiencies of existing architectures, we first provide a possible interpretation of graph diffusion patterns from the perspective of control. Similar to the physical process of thermal diffusion, the information contained in nodes can also be passed under the control of graph topology. To promote the flexibility of the model and alleviate over-smoothing phenomenon, we further propose a graph diffusion neural network controlled by the State Space Model (namely GraphDiCtate). In particular, we devise a jointly learnable diffusion and controlling strategy that simultaneously controls natural diffusion tendencies and message passing patterns recognized from training data. Within a single controller, the update of latent states of nodes follows both structure-driven and data-driven principles. Experimental results show that the proposed model outperforms the baselines on various semi-supervised node classification benchmarks, and effectively alleviates the over-smoothing problem over different depths of layers compared with regular MPNNs, which enables future work to train very deep GNNs even Large Graph Models.
Link prediction is crucial for analyzing graph-structured data, with applications in diverse fields including recommendation systems, anomaly detection, social network analysis, time series analysis, and bioinformatics. Graph Auto-Encoders (GAEs) and Variational Graph Auto-Encoders (VGAEs) play a central role in this task by effectively predicting unknown or future connections between nodes through robust embedding representations. In this work, we conduct a comprehensive analysis of GAE and VGAE models, revealing significant correlations between the covariance matrix of node features and their embedding representations. We identify several limitations in existing models, such as redundancy in capturing central tendencies in non-central distributions, challenges in representing negative correlations, and difficulties in distinguishing nodes with similar features but different connectivity patterns. To overcome these challenges, we propose methodological enhancements that include preprocessing with mean subtraction to align more closely with the true data distribution, a bidirectional activation function to capture both positive and negative correlations more effectively, and a contrastive loss function to improve the discriminative power of the model. These improvements collectively refine the model’s ability to discern nodes with subtle feature differences. Extensive experimental validation on various datasets, both homophilous and heterophilous, demonstrates that our improvements significantly enhance performance over existing link prediction methods.
Efficient task allocation remains a critical challenge in mobile crowdsensing, especially for hybrid systems combining opportunistic and participatory users. Existing methods usually ignore the heterogeneity of task priorities, significantly compromising the completion of urgent tasks. To overcome this drawback, an urgency-aware two-stage task assignment framework (UTTA) is proposed. The problem is formulated as an overall utility maximization model based on a task urgency-sensitive utility metric. In the first stage, a gated recurrent unit-based trajectory predictor based on divided regions is adopted to identify opportunistic users capable of fulfilling tasks during their natural mobility in urban road network. Subsequently, the second stage designs a task packaging strategy that considers both urgency utility and spatial proximity, and employs an efficient allocation strategy to recruit appropriate participatory users for the urgent task packages, optimizing budget-constrained utility. Experiments on a real-world mobility dataset demonstrate that the proposed UTTA framework obtain superior performance in both overall utility maximization and urgent task completion compared to other baselines under the identical constraints.
A key challenge in deep learning is to address co-adaptation between some of the neurons. The current mainstream approach is to mitigate co-adaptation by randomly dropping some neurons. However, these simple random dropout methods based on Bernoulli distribution ignore the characteristics of the feature map and cannot effectively capture neurons with co-adaptation potential. As a result, they lead to innocent neurons being dropped incorrectly, severely reducing the dropout effectiveness and leading to additional computational overhead. In order to improve the effectiveness and efficiency of the existing dropout methods, this paper proposes a novel dropout method, Fuzzy Dropout, which utilizes Fuzzy C-Means (FCM) in the dropout layer. Fuzzy Dropout utilizes the feature map's attributes to drop out effective neurons that are more likely to lead to co-adaptation instead of dropping all neurons probabilistically, resulting in more efficient and effective mitigation of co-adaptation. By efficiently identifying and manipulating effective neurons, Fuzzy Dropout further enhances the model's generalization ability while reducing computational complexity. We conducted experiments on various deep neural network models using two different training approaches: training from scratch and fine-tuning. The results clearly show that Fuzzy Dropout outperforms other state-of-the-art dropout methods. It demonstrates superior generalization ability and usability in training from scratch and fine-tuning scenarios. At the same time, Fuzzy Dropout reduces computational complexity by explicitly identifying and targeting data more likely to cause co-adaptation phenomena. Our experimental results show that Fuzzy Dropout can improve the test set accuracy by up to 2.08% while reducing the computational complexity by nearly 95% compared to the existing Dropout methods.
The whole process simulation model of municipal solid waste incineration (MSWI) with fixed parameter setting values makes it difficult to accurately map the combustion state under different working conditions. As a result, the error between the simulation and the actual results is large when the model simulates different working conditions. Therefore, this paper proposes a correction method for the whole process simulation model parameter based on the MSWI plant actual data under digital twin platform. First, the MSWI whole process numerical simulation model is constructed, and a parameter setting values correction strategy is established. Next, by analyzing the reasons for the errors between the simulation results and the actual results in the model, the parameters that need to be corrected are determined. Then, the parameter correction problem is converted into an optimization problem, and the parameter setting values are optimized by an adaptive genetic algorithm (AGA) combined with actual data. Finally, the accuracy of the whole process simulation model and the effectiveness of the proposed parameter value setting correction method are verified by the actual data of the plant. Under the support of the MSWI digital twin platform, the results show that when the grate speed increases, the temperature correction value generally shows a first decreasing and then increasing trend. The fluctuation range of parameter adjustment is −7.55%∼10.11%. When the proportion of primary air gradually increases, the temperature correction value shows a trend of first decreasing and then increasing. The fluctuation range of parameter adjustment is −16.86%∼3.43%.
The generation process of diffusion models is often complex and slow due to the numerous iterative steps and the high dimensionality of token sequences involved. To address the substantial inference latency and computational overhead associated with diffusion models, several token-reduction methods have emerged. However, existing token reduction methods lack effective evaluation criteria and practical architectural design, hindering the screening of important tokens. This paper introduces a cross-modal token reduction (CTR) method that preserves and optimizes important tokens to accelerate diffusion model inference and reduce computational burden. The CTR method employs token-level cross-modal contrastive learning (TCCL) to align image tokens with the global text condition in a shared semantic space. A semantically aware importance metric (SAIM) then quantifies the contribution of tokens during generation. The paper further presents a spatial-semantic token reduction (SSTR) method that combines spatial and semantic information to prune redundant tokens early and merge less important tokens later, thereby significantly reducing computational load while maintaining generation quality. Importantly, CTR requires no retraining or fine-tuning of the diffusion model backbone network. Only lightweight projection layers are trained offline and then used as plug-andplay modules during inference. Experiments on the COCO30 K dataset demonstrate that CTR achieves a 1.89-fold speedup on Stable Diffusion v1.5 and a 1.60-fold speedup on Stable Diffusion v2.1, with the Fréchet inception distance (FID) decreasing by 1.08 and 2.14.
Consensus control of nonlinear open multi-agent systems (OMAS), such as robotic swarms and smart grids, is vital for ensuring collective coordination in dynamic environments where agents frequently join or leave the network. Focusing on a class of nonlinear OMAS on time scales, we define each opening instant as a sampling instant and propose a physics-informed reinforcement learning algorithm (PIRL) to address the open topology for nonlinear OMAS. Initially, an impulsive transition process transforms open dynamic topologies into static configurations while preserving agent information and ensuring minimal oscillation during the transition process. Furthermore, the PIRL algorithm leverages physics constraints to enhance parameter convergence speed and adaptability in OMAS under open topologies. Finally, this advanced PIRL algorithm, tailored for nonlinear OMAS, addresses complex behaviors, non-convexity, and coupled interactions. Simulation comparisons demonstrate superior consensus control accuracy, reducing the average tracking error by 59.50%, 82.25% and 99.73% compared with multi-agent deep deterministic policy gradient (MADDPG), distributed sampled-data control, and model predictive control-reinforcement learning (RL-MPC) based algorithm, respectively. Real-world UAV experiments further validate the practical effectiveness of the proposed PIRL algorithm.
The online service platform offers a comprehensive repository of publicly available data for multi-criteria decision-making research. However, the issue of data quality has emerged as a significant factor impeding the reliability of decision-making processes. To address this concern under data-driven conditions, the theory of basic uncertain information provides a solution by simultaneously characterizing information and its credibility. In this study, we focus on information distribution modeling and information quality modeling as starting points. We integrate the probability distribution function-based collective intelligence knowledge distribution aggregation model with the basic uncertain linguistic information (BULI) model to propose a decision framework based on BULI-joint probability distribution function (BULI-JPDF). Firstly, we extend the one-dimensional quantile aggregation (QA) method to accommodate two-dimensional basic uncertain linguistic information and introduce a two-dimensional quantile aggregation (TDQA) method. Our verification demonstrates that TDQA maintains key properties of QA: after TDQA aggregation, the expectation value of JPDF equals a linear weighted combination of initial JPDF means; variance is smaller than a linear weighted combination of initial JPDF variances; and sharpness after TDQA aggregation surpasses that achieved through two-dimensional linear combination aggregation. Secondly, within the BULI-JPDF decision framework, we define three types of measures for consensus level, confidence level, and fairness level. Subsequently, we construct a weight optimization model aimed at maximizing these three measure levels as objectives. Finally, we apply the proposed TDQA aggregation theory and multi-objective weight optimization model to address car product ranking problems. We present a novel product ranking approach based on the BULI-JPDF decision framework and substantiate its scientific rigor and reliability through comprehensive case analysis.
Existing robust multi-objective evolutionary algorithms (RMOEAs) typically apply a uniform optimization strategy to all individuals for searching robust optimal solutions, overlooking the sensitivity differences of decision variables to perturbations, which may not only result in inefficient use of computational resources but can also compromise the search efficiency and overall optimization performance of the algorithm. Additionally, balancing optimality and robustness in robust multi-objective optimization remains a key challenge. To address these issues, we propose a robust multi objective evolutionary algorithm based on the Wasserstein distance and the R2 indicator, termed RMOEA/WR. The algorithm consists of two key components. First, a novel decision variable classification strategy based on Wasserstein distance categorizes variables into low-robustness and high-robustness groups. Then, optimization is performed in two stages: for low-robustness decision variables, the MOMBI-II algorithm, based on the R2 indicator, is applied; for high-robustness decision variables, dominance and perturbation effects are considered, where dominant individuals before and after perturbation are retained, and non-dominant individuals before perturbation form a new population, followed by optimization using the R2 indicator. Extensive experiments on benchmark problems and a real-world car side-impact scenario demonstrate that RMOEA/WR outperforms representative algorithms in robustness and convergence. Statistical analysis further confirms its superiority in obtaining high-quality robust solutions.
Detecting electricity theft is essential for promptly identifying and preventing unauthorized use of electricity, which helps in mitigating economic losses and ensuring the secure and stable operation of power systems. However, existing centralized machine learning-based detection methods still face several challenges: 1) high communication costs due to centralized data transmission, 2) significant privacy leakage risks arising from the need to aggregate sensitive user electricity data, 3) suboptimal detection accuracy caused by insufficient utilization of spatio-temporal characteristics of electricity consumption data, and 4) limited data sharing resulting from data privacy. To simultaneously address these four challenges, this study introduces a pioneering privacy-preserving framework for distributed electricity theft detection, that enjoys two-fold ideas: 1) it develops a theft detection model leveraging a Spatio-temporal Attention Fusion Network (STAFN) to enhance detection accuracy, and 2) It designs an efficient federated learning (FL) framework for model training, which enables distributed model training without aggregating raw user electricity data, directly tackling the problem of limited data sharing while reducing data transmission volume to lower communication costs. Furthermore, the framework integrates the modular additive one-time pad (MOTP) protocol for secure model aggregation, effectively mitigating privacy leakage risks during the federated training process. Experiments conducted on real-world datasets demonstrate that the proposed STAFN model outperforms existing state-of-the-art models in both centralized and federated learning settings, under independent and identically distributed (IID) and non-independent and identically distributed (Non-IID) data conditions. Meanwhile, the proposed framework achieves superior performance in terms of privacy preservation, communication efficiency, and data sharing feasibility compared to traditional centralized methods.
Driver fatigue detection is critical for road safety, as physiological signals provide objective evidence of a driver's cognitive state. While deep learning offers promising solutions, existing methods face two primary challenges: unstable multi-modal fusion caused by variable signal quality, and high computational costs that hinder real-time deployment. To this end, we propose a lightweight teacher-student framework, termed PCA-KD, for driver fatigue detection. First, we introduce Prompt-Conditioned Attention (PCA), a bottleneck fusion module that dynamically regulates cross-modal information exchange and suppresses modality-specific noise via conditional prompts. Second, to ensure real-time efficiency without sacrificing performance, we design a Multi-Level Knowledge Distillation (KD) strategy. This strategy effectively transfers both robust structural fusion priors and soft predictions from a high-capacity teacher to a compact student network. Extensive experiments demonstrate that the proposed PCA-KD achieves state-of-the-art detection accuracy while significantly reducing computational overhead, confirming its high suitability for real-time deployment on resource-constrained vehicular devices.
Traffic flow prediction is essential for intelligent transportation systems. Existing deep learning models often face limitations. Transformer-based methods primarily focus on temporal dependencies while overlooking spatial correlations, and Graph Neural Networks struggle with dynamic temporal variations. To address these issues, we propose an innovative framework named Ensembled R-CrossNET. First, a distribution-aware time-segment embedding method is designed to mitigate distribution shifts in non-stationary traffic data. Next, a novel two-stage self-attention mechanism is introduced to simultaneously capture long-range temporal dependencies and cross-dimensional dependencies, which serve as implicit spatial correlations within time segments, while balancing representational capacity. Finally, a Differential Evolution algorithm is employed to integrate multiple base models, significantly improving prediction robustness. Experiments on five real-world datasets show that Ensembled R-CrossNET outperforms existing methods, demonstrating the effectiveness of our joint temporal and cross-dimensional modeling and ensemble learning strategies.
Representation and retrieving meaningful actions and activities from surveillance video is a challenging task. It requires the extraction of semantic information efficiently from surveillance video, which is challenging due to the disparate distribution of the activity over the frames. This is due to the presence of numerous irrelevant or redundant frames that can obscure significant actions, making it challenging to accurately identify and classify activities. To address this issue, the proposed work presents a temporal knowledge graph and a keyframe selection process based on generated captions to store only the most relevant frames from the video. Moreover, an interactive knowledge graph has been added to store frames as images for future retrieval and for searching events and actions within the videos. Target objects are detected and tracked throughout for improved surveillance. The proposed approach is evaluated on the UCF-Crime, VIRAT and THUMOS14 datasets, which contain scalable and complex events. The framework achieves 69.8% accuracy in the mAP@tIoU(0.1–0.5) metric. Ablation experiments were conducted to evaluate the effectiveness of the proposed method in accurately representing the knowledge embedded within surveillance videos.
