Table of Content

  

20 March 2026
Volume 61 Issue 3

Encrypted traffic detection based on path signature features representation learning

JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE). 2026, 61(3):  1-10.  doi:10.6040/j.issn.1671-9352.9.2025.002

Abstract ( 16 )   PDF (4871KB) ( 7 )   Save

References | Related Articles | Metrics

Aiming at the problems of insufficient extraction of interactive behavioral features between encrypted flows, a PSFREL(Path Signature Feature Representation Learning)based encrypted flow detection method is proposed.Signature feature representation learning(PSFREL), which uses path signatures to characterize the hidden, unaffected by encryption interactions between traffic flows, uses an autoencoder to extract local features at the field level, and uses the residual network Cam-resnet, which combines the attention mechanism of the channel, to extract the global features of the traffic flow, forming a multi-granularity flow features for encrypted traffic detection. Comprehensive benchmarking across four encrypted network flow datasets(e.g., ISCX VPN-nonVPN)showcases the PSFREL frameworks capability to attain a 94.91% mean F1-Score.

Forensic analysis of poster design infringement based on visual salient features

YANG Bin, SUN Jiannan, CAO Enguo, LI Zichuan, ZHOU Zhili

JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE). 2026, 61(3):  11-19.  doi:10.6040/j.issn.1671-9352.9.2025.003

Abstract ( 13 )   PDF (18101KB) ( 3 )   Save

References | Related Articles | Metrics

Traditional clone detection methods primarily rely on pixel-level image similarities, often overlooking conceptual similarities in core design elements, particularly in compositional layouts. To address this limitation, we propose a forensic method for detecting poster design infringement based on visual saliency features, aimed at assisting experts in identifying and assessing conceptual plagiarism. To achieve this goal, a sophisticated deep learning model comprising four sub-networks is developed to process complex visual elements in design works and explicitly delineate key layout structural relationships. By computing conceptual feature similarities between posters and existing works, proposed method effectively identifies designers infringing behaviors. The experimental results demonstrate significant improvements in accuracy compared to traditional approaches in poster design infringement forensic analysis.

Chaincode vulnerability detection method based on pre-training model

JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE). 2026, 61(3):  20-28.  doi:10.6040/j.issn.1671-9352.9.2025.001

Abstract ( 12 )   PDF (3169KB) ( 5 )   Save

References | Related Articles | Metrics

Aiming at the problem of security vulnerabilities in chain codes in the consortium chain Hyperledger Fabric, a deep learning vulnerability detection network based on vulnerability subtrees and pre-trained models is proposed. The detection method includes two key stages: first, the chain code is extracted into an abstract syntax tree through an automated tool, and a vulnerability subtree structure VB-tree is designed to ensure that the model focuses on key vulnerability features. On this basis, it is converted into a data flow graph based on the data and control dependencies between program statements; second, the extracted features are processed using a pre-trained model to accurately identify potential vulnerabilities. Finally, chain codes of 6 935 open source projects in different fields are collected from Github to construct a dataset that can be used to evaluate the effectiveness of the method. Experimental results show that when detecting 21 types of vulnerabilities in chain codes, the average F1 score of the model is 93.68%, which is better than existing methods.

Secure weighted aggregation for VFL with malicious passive parties

ZHANG Zhengyin, WANG Lingling, HUANG Mei, ZHANG Yuxing, SONG Jiaorong

JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE). 2026, 61(3):  29-43.  doi:10.6040/j.issn.1671-9352.9.2025.004

Abstract ( 12 )   PDF (8329KB) ( 4 )   Save

References | Related Articles | Metrics

Considering the problem that untrustworthy participants in vertical federated learning launch data poisoning attacks to hinder model training, and that semi-honest participants launch inference attacks to steal privacy information of other participants, a securely weighted aggregation scheme for vertical federated learning with malicious passive parties is proposed. First, a utility evaluation algorithm is combined to defend against data poisoning attacks, and the maximum tolerance distance is designed to filter the poisoned embedding vectors; Second, an adaptive weight calculation algorithm is designed to ensure that the model can still effectively resist data poisoning attacks and maintain high convergence rate and accuracy in long-tailed data scenarios. Finally, the masking mechanism and symmetric homomorphic encryption algorithm are utilized to protect the privacy of embedding vectors against privacy inference attacks. Theoretical analysis and simulation results show that the proposed protocols has better computational efficiency and model performance, can effectively resist privacy inference attacks and data poisoning attacks, and improves the model accuracy by about 5%-10% compared with the latest related work.

Hate speech detection based on pre-trained models

JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE). 2026, 61(3):  44-53.  doi:10.6040/j.issn.1671-9352.1.2024.044

Abstract ( 13 )   PDF (1944KB) ( 6 )   Save

Related Articles | Metrics

To accurately detect and identify hate speech, the dataset samples are expanded and balanced by fine-tuning the large language model. The RoBERTa-Attention-GRU-TextCNN model is constructed based on the pre-training model RoBERTa, leveraging the powerful feature capture and extraction capabilities of deep learning for the analysis and mining of text sequence data. Firstly, the RoBERTa model is used to extract features from the text data; then, the self-attention mechanism is used to obtain the dependencies between words; finally, the acquired feature matrix is input into the GRU-TextCNN layer to capture deeper semantic information and local features. Two publicly available datasets provided by TweetEval are used to evaluate the model effect, and the experimental results show that the model has a better detection effect compared to the traditional hate speech detection model.

Attribute enhanced temporary group recommendation algorithm fusing long-and short-term interests

WANG Zhixuan, PANG Jifang, WANG Zhiqiang, SONG Peng, LI Ru

JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE). 2026, 61(3):  54-65.  doi:10.6040/j.issn.1671-9352.1.2024.040

Abstract ( 13 )   PDF (4338KB) ( 5 )   Save

References | Related Articles | Metrics

Group recommendation aims to provide recommendation services for group users, with the ultimate goal of meeting the different preference needs of group members. Most existing group recommendation algorithms are designed for fixed groups, ignoring a large number of temporary groups with randomness and specificity. In order to further expand the application scenarios of group recommendation algorithms and effectively address the shortage of historical interaction information in temporary groups, attribute enhanced temporary group recommendation algorithm fusing long-and short-term interests(ALSTG)is proposed. Firstly, item attributes information are injected into the overall historical interactions between users and items and the users short-term interaction sequences, respectively. Then, users long-and short-term interests are learned by comprehensively using hypergraph networks, graph neural networks, and gated recurrent units. Furthermore, the long-term interests of members are aggregated into the long-term interests of group through attention mechanism. At the same time, a similarity measurement strategy between the short-term interests of members and the long-term interests of the group is designed to calculate the weights of members, and the short-term interests of the group are obtained through weighted fusion. On this basis, a long-and short-term interest contrastive learning approach is adopted to maximize the consistency between the two types of group interests. The model is jointly optimized through comparative loss and recommendation loss to obtain high-quality group comprehensive representation, thereby achieving accurate recommendation for temporary groups. Finally, the feasibility and effectiveness of the proposed model were verified through comparative analysis and ablation experiments on two real datasets. The experimental results show that item attribute information and user short-term interests can effectively enhance the quality of temporary group representation and significantly improve the recommendation performance of the model.

Method for verbose queries reduction by integrating key and latent concepts

ZHU Mingyang, HUANG Yuxin, YU Zhengtao

JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE). 2026, 61(3):  66-74.  doi:10.6040/j.issn.1671-9352.1.2024.061

Abstract ( 9 )   PDF (2186KB) ( 4 )   Save

References | Related Articles | Metrics

Query reduction aims to enhance retrieval recall and precision by simplifying and condensing lengthy queries while retaining key information. Traditional methods often rely on statistical approaches or pre-trained models to extract keywords from lengthy queries for retrieval input. However, these methods struggle with query complexity(e.g., synonym and polyseme)and often lose crucial information. To address these issues, a method integrating key concepts and latent concepts for verbose query reduction is proposed. This approach integrates key concepts representing the core content of the query with latent concepts crucial for query understanding but not explicitly expressed to generate more comprehensive and effective queries. Specifically, pre-trained models generate concise and effective queries as key concepts, while pseudo-relevance feedback methods extract latent concepts from relevant document sets of the original query. Finally, both are combined to form the query reduction for improved retrieval. Experimental results on the Robust2004 dataset using a dense retrieval model show that the proposed method improves R@1000 and NDCG@10 by 2.1% and 3.6%, respectively, compared to baseline models.

Prototype-based recommendation method with uniformity constraints and importance balancing

CAO Yuxiang, LIAN Tao, WANG Long, JING Xingbo, DOU Haocheng

JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE). 2026, 61(3):  75-85.  doi:10.6040/j.issn.1671-9352.2.2024.089

Abstract ( 14 )   PDF (1945KB) ( 3 )   Save

References | Related Articles | Metrics

Prototype-based recommendation algorithms can achieve explainable recommendations by learning a set of user prototypes(or item prototypes)that represent typical preferences or common characteristics, as well as the association strength between users(or items)and prototypes. However, existing algorithms overlook the differences between prototypes and the load balancing among them, and hence cannot fully release the expressive power of the model. Therefore, a prototype-based recommendation method ProtoMF++ with uniformity constraints and importance balancing is developed on top of ProtoMF. This method added uniformity constraints between user prototypes(or item prototypes)and enhanced the differences between prototypes by minimizing the logarithm of the average pairwise Gaussian potential between prototype representations. In addition, the load importance of each prototype is defined as the total association strength between it and all users(or items), and the coefficient of variation of their load importance is minimized to realize importance balancing across different prototypes. Experiments are conducted on three benchmark datasets, and the results show that ProtoMF++ outperforms existing prototype-based recommendation methods. For example, on the Baby dataset, the values of HitRatio@10 and NDCG@10 increase by 4.74% and 10.64%, respectively.

Center moment discrepancy multimodal sentiment analysis based on self-attention mechanism

CHEN Zhongyuan, LU Chong

JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE). 2026, 61(3):  86-95.  doi:10.6040/j.issn.1671-9352.0.2024.230

Abstract ( 16 )   PDF (1790KB) ( 4 )   Save

References | Related Articles | Metrics

A center moment discrepancy multimodal sentiment analysis based on self-attention mechanism(SA-CMD)is proposed, aiming to address issues related to modality correlation mining, feature fusion strategies, and label updating mechanisms in existing models. First, an encoder is used to encode the extracted feature sequences, and the weights of each modalitys features are dynamically adjusted through a self-attention mechanism to capture the complex dependencies between modalities. Next, the center moment discrepancy method is introduced to dynamically optimize feature representations and label distributions, enhancing the models robustness. During the feature fusion process, the model calculates the distance discrepancy between modality features and their respective positive and negative centers to generate more accurate feature labels, further improving the quality of the fused features. Finally, a linear layer is used to project the fused features onto a lower-dimensional space for prediction. Experimental results show that SA-CMD outperforms existing baseline models in the public CMU-MOSI and CMU-MOSEI datasets across various evaluation metrics, especially in terms of the Pearson correlation coefficient, binary classification accuracy, and seven-class classification accuracy. Ablation experiments further verify the key role of the self-attention mechanism and the center moment discrepancy method in enhancing model performance, fully demonstrating the effectiveness and robustness of the SA-CMD model in multimodal sentiment analysis tasks.

Dynamic event-triggered practical fixed-time consensus for second-order linear multi-agent systems

WANG Chuanhao, LI Zonggang, NING Xiaogang, CHEN Yinjuan

JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE). 2026, 61(3):  111-123.  doi:10.6040/j.issn.1671-9352.0.2024.029

Abstract ( 14 )   PDF (4631KB) ( 8 )   Save

References | Related Articles | Metrics

Aiming at the problem that the second-order linear multi-agent systems event-triggered actual fixed-time consensus mostly adopts static triggering conditions, with too many triggering times and high system energy consumption, two dynamic event-triggered actual fixed-time consensus control protocols are proposed. Based on the controller with tracking error and hyperbolic tangent function, a continuous communication consensus control protocol is proposed. The internal dynamic variables adjusted by the relative state of the agent in real time are introduced in the event triggering condition, and the trigger threshold of the agent is adjusted in real time. The intermittent communication consensus control protocol uses the information of the agent trigger time to avoid continuous communication between agents. It is verified that under the two control protocols, the system can achieve the actual fixed-time consistency, and avoid the problems that the convergence time is limited by the initial state of the agent and the Zeno behavior. The simulation results show that compared with the existing static event-triggered scheme, the proposed dynamic event-triggered scheme reduces the number of triggers of the agent, thereby reducing the energy loss of the system, and is more suitable for the actual system with limited communication computing resources.

Intuitionistic fuzzy locality preserving projection least squares twin support vector clustering

WANG Shunxia, HUANG Chengquan, CAI Jianghai, YANG Guiyan, LUO Senyan, ZHOU Lihua

JOURNAL OF SHANDONG UNIVERSITY(NATURAL SCIENCE). 2026, 61(3):  124-134.  doi:10.6040/j.issn.1671-9352.0.2024.118

Abstract ( 18 )   PDF (5457KB) ( 11 )   Save

References | Related Articles | Metrics

To solve the problem that the local structure information of data samples is not fully utilized and the sensitivity of the algorithm to noise leads to the decline of the clustering effect, this paper proposes the intuitionistic fuzzy local preserving projection least square twin support vector clustering method. Fuzzy scores are assigned based on the distance between samples and centroids and the heterogeneity of the samples, given weights to the samples, and the local geometric structure information of the training sample is fully utilized to provide prior information about the sample neighborhood, which not only reduces the influence of noise and outliers on the performance of the algorithm, but also effectively solves the clustering problem of data. Experiments are conducted on several datasets, and the significance of the proposed algorithm is verified by statistical analysis. These experimental results demonstrate that the proposed algorithm has better robustness and clustering performance than other existing algorithms.