Fang Zuo

Long videos contain substantial temporal redundancy, whereas answer-relevant evidence is often sparse, distributed, and unevenly sampled. Under a limited visual budget, uniform sampling may miss brief events and global similarity ranking may over-select redundant observations. We propose Query-Conditioned Utility Profiling and Allocation (QUPA), a lightweight front-end that estimates a continuous, query-conditioned utility profile from complementary visual cues and converts it into a compact visual input for a frozen multimodal large language model. Contrastive Distribution Knapsack (CDK) partitions the temporal axis at profile transitions, allocates regional frame quotas using probability mass and saliency contrast, suppresses local redundancy, and assigns spatial resolutions according to relative utility. The core contribution is the coupling between utility profiling and distribution-aware allocation: the profiling stage provides a structured temporal signal, and CDK translates it into a budget-constrained selection. QUPA is trained with continuous pseudo-utility targets generated by a frozen teacher model (Qwen2-VL-7B), which is removed at inference. On LongVideoBench, Video-MME, MLVU, and NExT-QA, QUPA yields improvements over the corresponding frozen baselines under matched frame budgets, with a maximum gain of 6.2 percentage points on LongVideoBench when coupled with Qwen2-VL at 32 frames. Controlled comparisons further characterize its accuracy, visual-token, and selector-latency trade-offs.

Currently, research on graph neural networks (GNNs) has achieved certain advancements. However, most studies are conducted under the assumption that data labels are correct. In reality, numerous unexpected factors can introduce noise into the data, leading to a decline in the performance of GNNs. To address this issue, this manuscript proposes a new robust graph neural network method called Dual-Way Diffusion and Aggregation (DWDA) for node classification with noisy labels. Specifically, we first train a GCN model and save its parameters as a teacher model to generate soft labels. Then, using both soft labels and hard labels, we perform data graph diffusion followed by training the student model. The graph diffusion technique enables the smooth propagation of global information and the construction of node affinity, thereby enhancing the model's ability to capture global information and effectively refine the relationships between nodes. The soft labels generated by knowledge-distillation provide rich information, helping the model to better understand the data distribution and improve node classification accuracy. The comprehensive experimental results show that our method significantly outperforms existing state-of-the-art methods in node classification tasks with different noisy label ratios, demonstrating the effectiveness of DWDA in classification tasks with varying levels of noisy labels.

Wireless Sensor Networks (WSNs) consist of distributed sensor nodes that collect and transmit environmental data, often in resource-constrained and unsecured environments. These characteristics make WSNs highly vulnerable to various security threats. To address this, the objective of this research is to design and evaluate a deep learning-based Intrusion Detection System (IDS) that is both accurate and efficient for real-time threat detection in WSNs. This study proposes a hybrid IDS model combining one-dimensional Convolutional Neural Networks (Conv1Ds), Gated Recurrent Units (GRUs), and Self-Attention mechanisms. A Conv1D extracts spatial features from network traffic, GRU captures temporal dependencies, and Self-Attention emphasizes critical sequence components, collectively enhancing detection of subtle and complex intrusion patterns. The model was evaluated using the WSN-DS dataset and demonstrated superior performance compared to traditional machine learning and simpler deep learning models. It achieved an accuracy of 98.6%, precision of 98.63%, recall of 98.6%, F1-score of 98.6%, and an ROC-AUC of 0.9994, indicating strong predictive capability even with imbalanced data. In addition to centralized training, the model was tested under cooperative, node-based learning conditions, where each node independently detects anomalies and contributes to a collective decision-making framework. This distributed approach improves detection efficiency and robustness. The proposed IDS offers a scalable and resilient solution tailored to the unique challenges of WSN security.

Autonomous underwater vehicles (AUVs) perform a wide range of functions, since underwater circumstances are diverse and varied and resources are plentiful. However, existing prevalent theoretical computation methods for classifying underwater environments are unable to keep up with the constantly evolving and complicated undersea world. Therefore, in the design process of multi-AUV cooperative systems, there is usually some uncertainty in the parameters. This uncertainty creates some challenges in the design process and affects the system’s performance. In this study, a reliability-based multidisciplinary design optimization is presented, in which some of the problem parameters are uncertain. In this regard, it is assumed that some of the problem parameters are in the form of fuzzy numbers. We develop a multi-AUV cooperative system dependability model based on fuzzy control, leveraging the membership function of fuzzy control to thoroughly evaluate the environmental effects. We propose a method to evaluate the reliability of multi-AUV cooperative systems, which is of practical value. This solution can be used to evaluate and examine the cooperative system’s dependability in an unidentified underwater environment. The factors contributing to the reliability of multi-AUV cooperative systems are obtained. The results show that these contributing factors reflect different aspects of reliability for multi-AUV cooperative systems. It is possible to determine the reliability variation in the population of underwater vehicles operating in a complex underwater environment, which establishes a solid foundation for the further optimization of multi-AUV cooperative systems.

As combination therapy becomes more common in clinical applications, predicting adverse effects of combination medications is a challenging task. However, there are three limitations of the existing prediction models. First, they rely on a single view of the drug and cannot fully utilize multiview information, resulting in limited performance when capturing complex structures. Second, they ignore subgraph information at different scales, which limits the ability to model interactions between subgraphs. Third, there has been limited research on effectively integrating multiview features of molecules. Therefore, we propose ComNet, a deep learning model that improves the accuracy of side effect prediction by integrating multiview features of drugs. First, to capture diverse features of drugs, a multiview feature extraction module is proposed, which not only uses molecular fingerprints but also extracts semantic information on SMILES and spatial information on 3D conformations. Second, to enhance the modeling ability of complex structures, a multiscale subgraph fusion mechanism is proposed, which can fuse local and global graph structures of drugs. Finally, a multiview feature fusion mechanism is proposed, which uses an attention mechanism to adaptively adjust the weights of different views to achieve multiview data fusion. Experiments on several publicly available data sets show that ComNet performs better than existing methods in various complex scenarios, especially in cold-start scenarios. Ablation experiments show that each core structure in ComNet contributes to the overall performance. Further analysis shows that ComNet not only converges rapidly and has good generalization ability but also identifies different substructures in the molecule. Finally, a case study on a self-collected data set validates the superior performance of ComNet in practical applications.

Cerebral herniation is a life‐threatening neurological emergency, where timely and accurate prediction is crucial for improving patient prognosis. Due to its rapid imaging advantages, CT becomes the preferred choice for cerebral herniation screening. With the continuous development of artificial intelligence technology in the field of neurological diseases, CT‐based models provide significant support for computer‐aided clinical diagnosis. However, current research on cerebral herniation diagnosis remains limited. Existing methods rely on traditional machine learning or focus solely on midline shift detection, which not only exhibits strong subjectivity but also neglects key structures such as the brainstem and the rich information from sagittal CT images. To address these limitations, this study focuses on mid‐sagittal CT images including the brainstem and combines clinical data to construct a multimodal deep learning framework for cerebral herniation prediction. The model integrates mature and advanced deep learning architectures to extract and fuse features from CT images and clinical text data, employing multiscale convolution and attention mechanisms for diagnostic classification. The model is evaluated on datasets from two centers. Results show that on the internal test set, the model achieves accuracy, sensitivity, specificity, and AUC of 89%, 92%, 88%, and 0.94, respectively; on the external test set, it attains accuracy, sensitivity, specificity, and AUC of 81%, 82%, 80%, and 0.89, respectively, outperforming baseline methods and existing state‐of‐the‐art approaches. Additionally, when compared with radiologists on the internal test set, the model’s performance matches or exceeds the consensus of physicians. We also reveal the model’s focus region through visual analysis, which further deepens the understanding of the model’s prediction process and enhances its interpretability. Experiments demonstrate that the proposed method holds significant potential in assisting cerebral herniation diagnosis.

Securing task allocation is an enabling technology for Internet of Things (IoT) crowdsensing applications. It is a difficult problem to solve due to the complex multi-factor and multi-objective nature of IoT crowdsensing applications. Existing methods mainly focus on privacy protection for task allocation and secure task allocation, lacking an evaluation of security during the process of task allocation. To address the above issue, this paper proposes a Secure Task Allocation (STA) method based on the Analytic Hierarchy Process (AHP) and the Common Vulnerability Scoring System (CVSS). First, by considering task characteristics, security vulnerabilities, and participant information, a hierarchical structure is designed for secure task allocation with the AHP method. The optimal execution queue for the task is obtained by calculating task feature weights. Then, a quantitative evaluation method is designed to evaluate the security value of the tasks by using both the CVSS and AHP methods. The STA is obtained by a task-matching strategy based on task security values and participant information. The simulation results show that the STA method achieves high security and efficiency for task allocation and is superior to other comparative algorithms.

In recent years, large language models (LLMs) have undergone rapid development, demonstrating significant potential across various domains. As Internet technology advances, Internet-connected devices find increasingly meaningful applications, including device type and owner identification, necessitating improved performance. This paper aims to explore the utilization of LLMs in recognizing cyberspace device attributes through fine-tuning with instruction templates. These templates enable learning from device attribute data and uncovering potential associations, facilitating efficient and accurate recognition of cyberspace device attributes. Experimentation confirms the method's effectiveness and performance, achieving high-precision recognition of device attributes in real network environments. Results show a notable 75.77% accuracy rate in device type identification and an impressive 96.90% accuracy rate in device owner identification. This research introduces novel ideas and methods for utilizing LLMs in network security, holding both theoretical and practical significance.

In recent years, with the rapid advancement of technology and information technology, reinforcement learning has demonstrated tremendous potential in complex networks. The interdependence among urban infrastructures is becoming increasingly significant, underscoring the need for rapid and accurate assessment of interdependencies among foundational urban industries. This paper aims to establish a network model based on input-output tables across multiple urban industries, and to develop a cascading propagation model using the RAS algorithm. By integrating the DDPG algorithm from reinforcement learning, we minimize costs to achieve loss targets in urban foundational industries. Additionally, we introduce an adaptive algorithm to optimize the execution steps and reward modules of the DDPG algorithm. Experimental results demonstrate that our proposed enhanced DDPG-SAM algorithm not only achieves faster convergence in single executions but also improves model accuracy.