This research proposes a secure, explainable, and context-aware governance framework for blockchain-based digital media contracts in multimodal artificial intelligence-enabled AIoT interactive systems. As digital licensing, NFT copyright management, royalty distribution, and cross-chain content circulation become increasingly embedded in smart media ecosystems, existing contract auditing approaches remain limited by unimodal analysis, weak explainability, black-box decision processes, and insufficient cross-platform generalization. To address these challenges, the proposed framework integrates cross-modal data fusion, graph neural networks, and evolutionary expert knowledge fusion to model smart contract code, abstract syntax trees, control-flow graphs, intercontract transactions, cross-chain records, licensing metadata, NFT copyright attributes, and multimodal IoT interaction logs as unified graph representations. By incorporating causal priors and differentiable rules, the framework supports transparent risk reasoning, verifiable explanations, and trustworthy decision support. The study contributes to explainable blockchain security, multimodal AI governance, and intelligent media systems by enabling more robust detection of copyright misuse, unauthorized licensing, abnormal content distribution, and cross-chain transaction risks.
Short-term traffic flow forecasting is essential for travel safety, congestion avoidance, and effective traffic management as an integral part of intelligent transport systems. Long short-term memory (LSTM) has become a promising technique for forecasting traffic flow. Unfortunately, the LSTM model does not achieve adequate forecast accuracy due to noisy traffic data and poor selection of hyperparameter optimization values. To address these limitations, this study proposes a composite SFFSDF-LSTM model that integrates sinusoidal firefly feature selection (SFFS) and dragonfly (DF) algorithms with an LSTM network. The SFFS algorithm performs adaptive feature selection, while the DF algorithm optimizes hyperparameters to improve convergence and prediction accuracy. The proposed model efficiently identifies the optimal set of LSTM features, weights, biases, and hyperparameters for domain traffic flow prediction while minimizing training errors. From the perspective of error analysis and predictive analytics, the predictive accuracy of the combined model is evaluated and found to be superior to traditional nonparametric and baseline deep learning models.
Wireless body area networks (WBANs) became one of the most pioneering technologies in the medical sector and are known for continuous health monitoring and real-time medical data transmission. However, holding a secure WBAN environment is crucial to protect health-sensitive data and patient privacy from various cyber threats. This article serves as a comprehensive survey of potential security risks, focusing on security in WBAN. A detailed taxonomy of possible attacks is introduced, categorizing threats based on fundamental security services: authenticity, integrity, confidentiality, availability, and nonrepudiation. To alleviate these challenges, an in-depth classification of countermeasures has been provided, portraying detection mechanisms and defense strategies. This article emphasizes the need to implement multilayered security frameworks, integrating strong encryption, authentication mechanisms, and advanced anomaly detection mechanisms to safeguard WBAN systems. Moreover, this study serves as a foundation for future research on enhancing resilience in WBAN technologies.
In the big data era, database management systems (DBMSs) are required to handle multiple data models (e.g., relational, document, graph, and key-value data) at the same time. Moreover, there is also an increasing demand for DBMSs to support the in-database execution of machine learning (ML) algorithms, commonly referred to as in-database ML. To meet these needs, many multi-model databases (MMDBs) that can handle different data models using a single storage engine, along with in-database ML integration, have been developed. However, it still lacks a unified benchmark to evaluate their performance on analytical query and in-database ML workloads. In this paper, we introduce M3Bench, a benchmark that fully tests the ability of DBMSs to perform analytical query and in-database ML tasks on multi-model data. M3Bench is built based on real-world data from an online education application. It consists of a flexible multi-model data generator, which can create relational, document, graph, and key-value data based on a scale factor and custom data distributions, and a suite of tasks that are representative analytical query and ML problems involving more than one data model. To validate the effectiveness of M3Bench, we benchmark four MMDBs, namely OrientDB, ArangoDB, Apache AGE, and AgensGraph, with the M3Bench workloads and highlight their advantages and disadvantages for each data model and task type.
Large Language Models (LLMs) have demonstrated strong generative capabilities in open-domain question answering (ODQA), but the lagging nature of their parameterized knowledge limits their ability to handle time-sensitive queries. Although Retrieval-Augmented Generation (RAG) mitigates this issue by introducing external knowledge, existing approaches primarily focus on optimizing retrieved passages, and their performance remains highly dependent on the quality of initial retrieval. To address this limitation, we propose the QAugment-Summary framework, which aims to enhance LLM performance in ODQA tasks through fine-grained evidence collection and deep integration. The framework first performs question augmentation and sub-question decomposition to enable more targeted information retrieval. Then, it adopts a collaborative retrieval-generation approach to collect candidate evidence for each sub-question. The core innovation lies in our multi-stage summarization mechanism: the system independently summarizes both the retrieved passages and the complementary passages generated by the LLM, then concatenates these summaries and conducts a second-stage deep summarization to form a highly condensed and information-rich explicit reasoning basis. This final summary effectively handles long-text context and filters irrelevant information, while also serving as a robust foundation for LLMs to generate the final answer. Experimental results on multiple ODQA benchmark datasets demonstrate a 9.5-point gain in Exact Match (EM) scores, indicating that QAugment-Summary substantially improves answer accuracy, information integration, and cross-document reasoning, thereby offering a promising direction for building more reliable LLM-based reasoning systems.
Scene Graph Generation (SGG) is crucial for semantic image understanding, supporting tasks like image-text retrieval and visual question answering. Traditional methods rely on extensive annotations, which limits the generalization. Weakly Supervised SGG (WSSGG) alleviates this burden by utilizing image captions; however, it often yields incomplete graphs, semantic simplification, and unreliable localization. We propose Boosting Scene Graph Completeness and Localization under the Weak Supervision (BSGCL) framework, which leverages Multimodal Large Language Models (MLLMs) to extract and align triples from images. BSGCL enhances relational diversity and semantic depth through context-aware reasoning and dynamic alignment, surpassing the limitations of caption-based supervision. Specifically, BSGCL comprises a relation evaluator that filters noisy triples and enhances localization, particularly in cases of ambiguity. Furthermore, we also use the in-context learning (ICL) strategy to enhance MLLM and LLM generalization, enabling adaptation without fine-tuning. Experiments on COCO and Visual Genome demonstrate that BSGCL significantly outperforms existing WSSGG methods, achieving higher Recall@K and meanRecall@K. These results show its effectiveness and scalability under weak supervision.
With the swift rise of online news recommendation platforms, people can access the latest news quickly and conveniently. News recommendation models are central to these systems, enabling the inference of user interests and the delivery of accurate suggestions. Many existing systems adopt supervised learning based on user behavior records. However, user feedback is typically noisy and insufficient, ultimately impairing the performance of news recommendation models. To alleviate this problem, in this paper, we propose PNRM, a news recommendation framework that integrates self-supervised contrastive pre-training. We explore intrinsic signals from the raw interaction data, alleviate the data sparsity problem, and obtain more robust representations from noisy user behaviors. PNRM follows a two-stage training paradigm comprising pre-training and fine-tuning. (1) In the pre-training stage, we design three contrastive learning tasks, namely user-user contrast, user-news contrast, and news-news contrast, to mine intrinsic supervisory signals from large-scale unlabeled interactions. (2) In the fine-tuning stage, the news encoder and the user encoder are initialized with the parameters from the pre-training stage and further fine-tuned with the click data. This two-stage pipeline first acquires robust representations using self-supervision, then fine-tunes them for recommendation, equipping the model with stronger generalization and task-specific performance. Experiments conducted on Microsoft News Dataset (MIND) substantiate that our proposed method outperforms the basic model of a strong BERT-enhanced two-tower baseline, which leads publicly available two-tower baselines on MIND, showcasing the effectiveness of our proposed method.
Traffic flow prediction is a fundamental task in intelligent transportation systems. A key challenge lies in effectively capturing the complex and dynamic spatio-temporal dependencies in traffic data. In recent years, Graph Neural Networks (GNN) and Transformer-based models have shown great promise in addressing this challenge. However, the inherent spatio-temporal heterogeneity of traffic flow data remains a significant obstacle. Existing models often struggle to model the rapidly changing spatio-temporal dependencies among road segments and differentiate the importance of extracted features at multiple spatial and temporal scales. To this end, we propose M2STFormer, a Mixture-of-Experts Multi-Level Spatio-Temporal Transformer (M2STFormer) for traffic flow prediction. Our model incorporates a Mixture-of-Experts (MoE) module to learn unique patterns of each spatio-temporal location, and a hierarchical selection mechanism to adaptively fuse multi-level spatio-temporal features. We validate the effectiveness of M2STFormer through extensive experiments on two public datasets and two newly collected real-world datasets. The results show that our method achieves state-of-the-art performance across multiple datasets compared to single-stage baselines. Furthermore, while greatly improving computational efficiency, our approach matches or surpasses the performance of representative two-stage pretraining methods on several key metrics.
The detection performance of autonomous vehicle object detection algorithms typically degrades significantly in long-tail scenes, potentially leading to serious accidents. Challenges such as complex backgrounds and substantial variations in object morphology further increase detection difficulty. To address these issues, we propose LTE-DETR, a novel object detector designed for long-tail scenes. Specifically, we design a Global-Local Collaborative Feature Adaptive Enhancement Module (GLAEM) to adaptively enhance features extracted by the backbone network, thereby improving the model’s focus on object-relevant features. Additionally, we design a Deformation-Aware Cross-Scale Feature Fusion Module (DA-CFFM) to enhance the model’s adaptability to diverse target morphologies by incorporating deformable feature learning. Compared with the baseline, our method shows superior performance in object detection tasks under long-tail scenes, which validates the effectiveness of LTE-DETR.
Influence maximization has been one hot topic in the field of network analysis in recent years. The essence of influence maximization is to find a given number, k, seed nodes, so that their influence can diffuse to as many other nodes as possible in the network. Considering the timeliness and the value attenuation of information in the transmission, we believe that information spreading along the shortest path from one node to another is the fastest way between the pair of nodes. The shortest paths originated from any node form a shortest-path tree, and using the shortest-path tree to estimate the influence of the root node is an effective approach. Motivated by this, we propose an influence maximization algorithm based on the shortest-path tree in this paper, we name it SPTIM (acronym for Shortest Path Tree based Influence Maximization). Firstly, for every node, the shortest paths are identified to form its shortest-path tree, then its influence is estimated according to the rules developed in this paper. Afterwards, the node with the largest estimated influence is selected as a seed. Then, To alleviate the problem of influence overlapping, we attenuate the one- and two-hop neighbors of the selected seed with different proportions considering the impact of community structure. Next, the node with the largest influence is selected as another seed accordingly. This procedure is repeated until the number of selected nodes reaches k. The experimental results on some real networks under the IC, LT, and SIR models show that the proposed method, SPTIM, is competitive in terms of performance and efficiency.
Time-series data offers significant potential for improving services but risks exposing sensitive information. Existing privacy mechanisms often distort critical temporal patterns, undermining data utility. To address this, we propose DP-ADSP, a dynamic pattern-aware data stream privacy protection mechanism that protects privacy while preserving essential time-series patterns. In particular, the core of our methodology lies in a dynamic pattern-aware importance sampling technique that intelligently determines whether to sample and perturb data points. To mitigate utility degradation from excessive perturbation, we develop an adaptive privacy budget allocation strategy that optimally balances privacy protection with data fidelity. Furthermore, we introduce an adaptive perturbation mechanism to enhance the privacy protection of data. We provide theoretical privacy guarantees and demonstrate DP-ADSP’s superiority over existing methods through extensive experiments on real-world datasets.
Heterogeneous Graph Neural Networks (HGNNs) have garnered significant attention in recent years. They are particularly effective at modeling complex structures and diverse node types in real-world data. However, effectively capturing both structural and semantic information in heterogeneous graphs while maintaining computational efficiency remains a critical challenge. In this work, we propose HTKD, a Heterogeneous Transformer with self-Knowledge Distillation framework. The model employs a graph structural encoder that captures topological relationships and node interactions, and a graph semantic encoder that extracts rich, context-aware semantics. These complementary embeddings are then adaptively fused via a Transformer to generate expressive node-level representations. To further enhance representation learning while reducing computational overhead, we incorporate a self-knowledge distillation mechanism, allowing deeper semantic insights to be distilled into shallower layers. Extensive experiments on multiple benchmark datasets demonstrate that our approach achieves superior performance. The code is publicly available at https://github.com/404LEO404/HTKD .
In scenarios where data is difficult to annotate and label noise is prevalent, traditional supervised learning methods are prone to performance degradation due to incomplete and noisy labels, which limits their classification capabilities. To address these challenges, we propose a semi-supervised classification framework based on abductive learning. By integrating knowledge graph reasoning with deep neural network modeling, the framework enhances the model’s ability to learn effectively in scenarios characterized by sparse annotations and noisy labels. The abductive learning framework generates high-quality label hypotheses through abductive reasoning, which are used both to complete unlabeled data and to correct erroneous annotations in the existing dataset, thereby improving the overall reliability of the training data. To prevent the model from over-relying on abductive labels, a hybrid loss strategy is proposed to balance and integrate supervision signals from both ground-truth annotations and abductive inferences. Experimental results demonstrate that the proposed method can effectively address issues such as label scarcity and noise, validating its capability to improve label quality and classification performance in semi-supervised learning scenarios.
Transformer-based methods have made impressive progress in low-level vision tasks, such as image super-resolution. However, we find that these networks can only utilize a limited spatial range of input information. Some networks enlarge the receptive field by expanding the window of self-attention leading to suffer from intensive computations. And how to ensure that the fine and natural texture details of the image are recovered while effectively reducing the model complexity to meet the demand of migrating the use on lightweight devices is a difficult problem in this field. In this paper, we present a lightweight facial image super-resolution reconstruction algorithm. It introduces the Pulse-Coupled Neural Network (PCNN) into attention mechanism thus making use of their complementary advantages of being able to utilize global statistics and strong local fitting capability. Moreover, to reduce the computational workload, the overall model adopts an generative adversarial network structure to incrementally generate images. We trained the model on the CelebA dataset and performed comprehensive performance and generalization tests on the remaining 200 images of the CelebA and Helen datasets. Experimental results show that our method achieves competitive performance while maintaining an lower parameter count when compared with previous leading methods, especially obtaining the highest LPIPS and MPS scores. Compared to methods with similar parameter magnitudes, our network shows significant improvements in all metrics, notably, a reduction in the FID by 45.9.
Graph neural networks (GNNs) have achieved remarkable success in unsupervised graph anomaly detection (GAD) tasks. The existing GNN-based unsupervised GAD models typically employ self-supervised learning to capture the intrinsic low-dimensional representations of data, thereby adapting to the inductive bias of GNNs toward homophily. However, they often ignore the anomaly-discriminative property of nodes, termed the one-class homophily property, i.e., normal nodes tend to have strong affinity with each other, while the homophily in anomalous nodes is significantly weaker than that in normal nodes. In this paper, motivated by the one-class homophily, we propose a novel local-affinity-based adaptive graph filter (LAF) to address label imbalance challenge in unsupervised GAD tasks. Our method generates subgraphs by removing heterophilic edges from the raw graph, thereby strengthening the isolation of anomalous nodes and enhancing high-frequency information of the graph structure. Subsequently, we apply an adaptive graph filter to each subgraph to dynamically integrate low-frequency and high-frequency information, so as to better model the anomalous features. Instead of minimizing the commonly used data reconstruction errors, our method optimizes the model by maximizing the local node affinity. The final result is obtained by averaging the anomalous scores from multiple models trained on different subgraphs, during the inference phase. Experimental results on six real-world GAD datasets show that LAF outperforms the existing baseline algorithms in unsupervised GAD tasks.
Multivariate time series data often suffer from missing values due to external factors such as sensor failures and data transmission errors. To address these challenges, this paper proposes a dual-branch spatiotemporal imputation model based on Transformer, named DBTF. The model captures both temporal and feature correlations through a dual-branch architecture. Specifically, the parallel temporal and feature paths are designed to extract long-range temporal dependencies and cross-feature spatial correlations, respectively. A cross-attention fusion module is introduced to enable adaptive interaction between the two paths. The feature branch leverages a patch-based representation combined with a feature attention mechanism to model complex inter-feature correlations. In parallel, the temporal branch employs a novel Projected Attention Embedding (PAE) module to enhance long-range temporal dependency modeling. The fusion module employs a gated cross-attention mechanism to dynamically adjust the importance weights of temporal and feature information. Finally, a multi-layer perceptron with residual connections is used as the prediction head to improve imputation accuracy. Experimental results on multiple real-world datasets demonstrate that DBTF can effectively complete missing data and exhibits strong robustness under various missing rates. This study provides an effective solution for accurate recovery of multivariate time series with complex missing patterns.
Cardinality estimation over sliding windows is a fundamental task in data stream analytics. While sketch-based methods offer efficient approximations with low memory usage, they inherently lack data privacy protection, making them vulnerable to inference attacks when shared across different parties or queried repeatedly. To address this issue, we propose SlideDP, a novel sketch that achieves accurate and efficient cardinality estimation under local differential privacy (LDP) in sliding window settings. SlideDP maintains a compact two-dimensional array, where each cell contains a 1-bit register and a small counter to jointly encode the presence and temporal validity of elements. Privacy is enforced via randomized response applied to each register, while a clock-based decay mechanism periodically decrements counters to track element expiration without requiring explicit timestamps. To reduce computational overhead, SlideDP incorporates ordered Poisson sampling for efficient updates. We further develop a concrete probabilistic model to support accurate estimation with theoretical guarantees. Experiments on both synthetic and real-world datasets show that SlideDP achieves accuracy comparable to non-private methods. In particular, when the window cardinality exceeds 2^15 , SlideDP maintains a relative error of approximately 5
Collaborative perception is crucial for the advancement of autonomous driving. However, collecting multi-agent collaborative data is extremely challenging and costly. To address this, researchers commonly utilize virtual datasets that are automatically labeled and can be generated in bulk to train and evaluate collaborative perception models. Despite this convenience, there are significant domain gaps between virtual and real-world datasets, which severely hinder model performance in real-world scenarios. To bridge this gap and unlock the potential of virtual data, we propose a virtual-real domain adaptation method for V2V collaborative detection based on Axially Focused Linear Attention, named AFLA-V2VDA, which comprises three core components. First, a Random Object Scaling (ROS) mechanism is introduced to dynamically adjust vehicle target sizes, mitigating the scale bias learned from the source domain. Second, a Spatially Adaptive Feature Domain Alignment (SFDA) module is proposed to reduce the feature distribution gap between virtual and real domains by leveraging spatial position awareness and dynamic feature weighting. Finally, a lightweight and robust feature fusion module (FuseAFLA) is designed to facilitate efficient cross-agent feature integration. Experimental results demonstrate that our domain adaptation approach significantly improves cross-domain collaborative perception performance.
Accurate classification and grading of data based on sensitivity levels are crucial for ensuring information security and compliance. Traditional methods often struggle with heterogeneous data sources and evolving regulatory standards. To address these challenges, we propose HMSC-FuseMaskNet, a novel neural network architecture designed for context-aware data security classification. The model integrates three key components: (1) a multi-source encoder that processes both structured and unstructured data; (2) a cross-source semantic fusion module employing attention mechanisms to align data representations with relevant security standards; and (3) a learnable feature masking layer that emphasizes critical features while suppressing irrelevant ones, enhancing interpretability. We evaluate HMSC-FuseMaskNet on real-world datasets from finance, healthcare, and education sectors. The experimental setup includes comparisons with baseline models such as traditional machine learning classifiers and existing deep learning approaches. Results demonstrate that our model achieves superior performance, with an average accuracy improvement of 4.7
Machine learning has shown great promise in advancing healthcare applications, particularly in predictive modeling for clinical outcomes. However, widespread adoption in medical practice remains limited due to critical concerns around explainability, reliability, and variability across institutions. To address these challenges, we apply a Causality-Informed Neural Network (CINN) framework and integrate expert-derived causal knowledge into the model architecture. By embedding causal relationships identified by physicians, CINN aims to enhance model transparency and generalizability across different clinical settings. In a case study focused on predicting the length of stay (LOS) in the Intensive Care Unit (ICU) for sepsis patients, we collaborate with medical experts to incorporate domain-specific causal structures. Experimental results highlight its superior explainability, robustness under distributional shifts, and improved transferability across cohorts. These findings suggest that CINN offers a promising pathway for building clinically aligned, actionable AI tools in healthcare.
