Computer Science

Survey of Communication Efficiency for Federated Learning

ZHENG Jianwen, LIU Bo, LIN Weiwei, XIE Jiachen

Computer Science. 2025, 52 (2): 1-7. doi:10.11896/jsjkx.240100023

Abstract

PDF(1727KB) ( 281 )

References | Related Articles | Metrics

As a distributed machine learning paradigm,federated learning(FL) aims to collaboratively train machine learning models on decentralized data sources while ensuring data privacy.However,in practical applications,FL faces the challenge of communication efficiency,as significant communication is required in each iteration to transmit model parameters and gradient updates,leading to communication costs far surpass computation costs.Thus,effectively enhancing communication efficiency poses a significant challenge in FL research.This paper mainly introduces the importance of communication efficiency in FL,and divides the existing research on FL communication efficiency into client selection,model compression,network topology reconstruction,and the combination of multiple technologies according to different emphases.On the basis of the existing research on FL communication efficiency,this paper summarizes the difficulties and challenges in communication efficiency in the development of FL,and explores the future research direction of FL communication efficiency.

Overview of Research on Post-quantum Cryptography Technology

WU Kun, HU Xiangang

Computer Science. 2025, 52 (2): 8-19. doi:10.11896/jsjkx.240500056

Abstract

PDF(1467KB) ( 527 )

References | Related Articles | Metrics

The development of quantum computing poses a significant security threat to classical cryptographic systems.Post-quantum cryptographic algorithms,which are theoretically capable of resisting quantum attacks,have become a hot topic of research at present.According to the classification of hardness assumptions,this paper first introduces the current state of research on post-quantum cryptographic algorithms such as lattice-based,code-based,multivariate,and hash-based,analyzing their technical characteristics and advantages and disadvantages.At the same time,combined with the results of the NIST post-quantum cryptography standardization,typical cryptographic algorithms of different technical routes are introduced.Finally,this paper summarizes the technical solutions for the current migration to post-quantum cryptography and proposes possible future development directions for post-quantum cryptography.

Coherent Legal Governance of Synthetic Data in AI Training

ZHANG Tao

Computer Science. 2025, 52 (2): 20-32. doi:10.11896/jsjkx.240900163

Abstract

PDF(1568KB) ( 179 )

References | Related Articles | Metrics

Artificial intelligence requires large,diverse,and high-quality data to train machine learning models,and collecting this real-world data can be very difficult and can threaten individual privacy,trigger bias or discrimination,and violate copyright.In practice,synthetic data,as an alternative solutionhas received widespread attention and is increasingly being used to train machine learning models.This paper explores the governance framework of synthetic data in AI training from the perspective of data jurisprudence,drawing on research from both data science and computer science.It first analyzes the logical premise of the importance of synthetic data in AI training from the normative level,i.e.,there is an obvious incompatibility between the protection of “small privacy” pursued by the personal information protection law and the demand for “big data” in AI training,which makes the deve-lopment of training data challenging,and the development of synthetic data for machine learning models challenging.The development of training data faces challenges,while existing legal and technological solutions suffer from ineffective governance.On this basis,the application scenarios and risk types of synthetic data in AI training are discussed.Finally,it is proposed to build a coherent legal governance framework for synthetic data in AI training from three aspects,guided by the “law 3.0 theory” and “data governance theory”:formulating rules for handling synthetic data,strengthening process governance of synthetic data,and developing assessment tools for synthetic data.

Performance Optimization of LSM-tree Based Key-Value Storage System Based on Fine-grained Cache andLearned Index

XU Ruida, LI Yongkun, XU Yinlong

Computer Science. 2025, 52 (2): 33-41. doi:10.11896/jsjkx.240200001

Abstract

PDF(2535KB) ( 229 )

References | Related Articles | Metrics

In the context of the big data era where the amount of data is growing rapidly,log-structured merge-tree-based(LSM-Tree-based) key-value storage systems are widely deployed in many NoSQL systems because of their excellent flexibility and scalability.However,when querying data in the traditional LSM-Tree based key-value storage system,the read amplification problem,which is caused by searching through multiple SSTables,generates additional I/O overhead.In this paper,we present FCLI-LSM,a new optimization design for key-value storage systems.FCLI-LSM improves the performance of LSM-Tree-based key-value storage by combining the optimization methods of fine-grained key-value pair caching and learned index.By analyzing data access hotspots.FCLI-LSM implements three-level classification of hot,warm,and cold key-value data.FCLI-LSM designs a fine-grained caching mechanism for hot data based on key-value separation,effectively reducing the additional I/O overhead caused by the read amplification problem.In addition,FCLI-LSM also designs a cache affinity optimization for learned index,which further improves the query performance of storage system for warm key-value data.Experimental results show that,compared with existing read optimization solutions,FCLI-LSM can reduce average read latency by more than 40% and increase system throughput by more than 1.7 times.

Study on Distributed Hybrid Storage Based on Erasure Coding and Replication

FU Xiong, SONG Zhaoyang, WANG Junchang, DENG Song

Computer Science. 2025, 52 (2): 42-47. doi:10.11896/jsjkx.231200021

Abstract

PDF(1889KB) ( 121 )

References | Related Articles | Metrics

With the rapid development of big data technology,cloud computing,computer technology and network technology,Internet data has shown explosive growth,and efficient storage of massive data has become an urgent challenge for current Internet technology.However,traditional multi-copy redundancy mechanisms result in huge storage costs,thus drawing attention to new storage solutions.In this context,a distributed hybrid storage strategy based on erasure coding and replica replication is proposed.Based on data characteristics,this strategy uses replica replication for hot data to ensure high reliability and performance,while erasure coding is used for cold data to improve storage utilization.Based on Newton's cooling law,the data files is divided into hot files and cold files,and an adaptive data temperature identification and hot and cold data adaptive dynamic allocation algorithm are introduced,so that the system can automatically adjust the ratio of hot and cold data at runtime,and then intelligently adjust the data storage strategy according to the the hot and cold conditions of real-time data,which reflects the system's adaptability in a dynamic environment.It not only enhances the system's adaptability to dynamic workloads,but also provides a new paradigm for the efficiency and flexibility of distributed storage systems in practical applications.This innovation has important promotion significance at both the academic and practical levels.At the same time,the effectiveness and usability of the strategy have been verified through simulation experiments,which provides new ideas for the optimization of distributed storage systems.

Study on Erasure Code Algorithm for Three Data Centers

SUN Jing, NIU Hongting, LIANG Songtao

Computer Science. 2025, 52 (2): 48-57. doi:10.11896/jsjkx.241000022

Abstract

PDF(2185KB) ( 122 )

References | Related Articles | Metrics

Erasure coding algorithms are widely applied in both single-data-center and multi-data-center environments.Current research on erasure coding algorithms primarily focuses on storage cost and repair bandwidth.However,challenges such as how to perform repairs across multiple data centers with limited dedicated bandwidth and switch constraints,and how to balance key factors like reliability and fault tolerance,have not been thoroughly analyzed or addressed.This study targets the three-data-center scenario,which is one of the most common multi-data-center configurations.First,it identifies four crucial factors for erasure coding in system design:redundancy,reliability,fault tolerance,and decoding bandwidth.Based on these factors,it then proposes ansingle-data-center local reconstruction code(S-LRC) algorithm that achieves optimal bandwidth repair within a single data center.Building on the S-LRC algorithm,it further develops the global local reconstruction code(G-LRC) algorithm to accommodate the architecture of a three-data-center setup.Compared to traditional coding schemes,the proposed G-LRC algorithm offers higher reliability,greater fault tolerance,and a lower decoding bandwidth penalty.Specifically,for two-node failures,the decoding bandwidth penalty of G-LRC is only 1/7～2/7 of that of traditional schemes.Finally,the G-LRC algorithm is implemented and validated in a large file storage system,where an optimal decoding decision algorithm is designed to reduce repair bandwidth,addressing the deployment challenges of non maximum distance separable(non-MDS) codes in practical systems.

Multi-view Multi-label Learning with Label Correlation Priors

SHENG Sirou, OUYANG Xiao, TAO Hong, HOU Chenping

Computer Science. 2025, 52 (2): 58-66. doi:10.11896/jsjkx.240600030

Abstract

PDF(2294KB) ( 145 )

References | Related Articles | Metrics

Multi-view multi-label learning,as a widely used machine learning and data mining technique in fields such as image classification,text mining,and bioinformatics,is receiving extensive attention from researchers.In this framework,samples are typically represented by multiple views and can be associated with multiple labels simultaneously.Although many methods have been proposed,many of them fail to fully integrate prior information to enhance learning performance,which often leads to unsa-tisfactory prediction performance.Aiming at this issue,this paper proposes a new multi-view multi-label learning method called multi-view multi-label learning with label correlation priors(MERIT).In the absence of labeled training data,this method acquires a multi-label prediction model by using only the prior of label correlations as weak supervision,thereby reducing the dependence on a large amount of annotated data.It not only adaptively adjusts the weights of different views but also accurately characterizes the similarity among samples of the same group by minimizing the discrepancy between sample similarity and label similarity.Experimental results on 8 multi-view multi-label datasets show that MERIT exhibits superior performance compared to similar methods.

Multivariate Time Series Forecasting Based on Temporal Dependency and Variable Interaction

WANG Huiqiang, HUANG Feihu, PENG Jian, JIANG Yuan, ZHANG Linghao

Computer Science. 2025, 52 (2): 67-79. doi:10.11896/jsjkx.240100167

Abstract

PDF(2939KB) ( 155 )

References | Related Articles | Metrics

Multivariate time series forecasting has a wide range of applications,such as power forecasting,weather forecasting.Although the latest models have achieved relatively good results,they still face the following challenges:1)it is difficult to fully consider the correlation between different variables in multivariate time series to make more accurate predictions;2)modelling the correlation between different variables usually requires a huge time and space cost.Current methods are mainly classified into va-riable-independent methods and variable-mixed methods.The variable-independent methods predict each variable based on its own information without considering the correlation between different variables;the variable-mixed methods embed all the variables into a high-dimensional hidden space without modelling the correlation between the variables in a targeted way,and cannot adequately capture the correlation between the variables.To address these challenges,this paper proposes a multivariate time series forecasting method FIID based on temporal dependence and variable interaction,which adequately models the correlations among different variables while greatly reducing the time and space costs.Specifically,this paper proposes variable fold based on the fact that correlations between different variables are usually sparse,which greatly reduces the time and space cost of subsequent mo-delling of correlations between different variables.Then this paper proposes the temporal dependence module to capture the global correlations within variables by linear transformation from the frequency perspective.Further,this paper defines the correlation between different variables as the correlation between different time periods of all variables,based on which this paper proposes the variable interaction module,which first aggregates the local information of the variables,and then models the global correlation between all the local features on this basis.With these two modules,not only the correlations between variables are adequately modeled,but also the time and space costs are greatly reduced compared to existing methods.The model FIID is experimented on twelve real datasets,and the results show that it achieves the best performance and possesses higher efficiency.

Graph Anomaly Detection Model Based on Personalized PageRank and Contrastive Learning

YUAN Ye, CHEN Ming, WU Anbiao, WANG Yishu

Computer Science. 2025, 52 (2): 80-90. doi:10.11896/jsjkx.240200005

Abstract

PDF(2872KB) ( 146 )

References | Related Articles | Metrics

Graph anomaly detection aims to detect abnormal nodes from attribute networks,and is highly valued by researchers due to its profound practical significance in many application fields such as finance,electronic trade,and spam sender detection.Traditional non deep learning methods can only capture the shallow structure of the graph,and researchers have proposed anomaly detection models based on deep neural networks to address this issue.However,these models do not take into account the centrality differences of nodes in the graph,which can lead to information loss or introduce noise from remote nodes when capturing local information of nodes.In addition,they ignore the feature information in the attribute space,which can provide additional anomaly monitoring signals.Therefore,this paper proposes a novel graph anomaly detection framework PC-GAD(personalized PageRank and contrastive learning based graph anomaly detection) from an unsupervised perspective.Firstly,a dynamic sampling strategy is proposed,which calculates the personalized PageRank vector of each node in the graph to determine the corresponding size of subgraph samples,avoiding the loss of local information and noise introduction.Secondly,for each node,the abnormal supervision signals are captured from the perspective of topology and attribute space,and the corresponding contrastive learning objective is designed to comprehensively learn potential abnormal patterns.Finally,after multiple rounds of contrast and prediction,the degree of abnormality of each node is evaluated according to the score of the output outlier.To verify the effectiveness of the proposed model,a large number of comparative experiments are conducted with the benchmark models on six real datasets.Experimental results have verified that PC-GAD can comprehensively identify abnormal nodes in the graph,and the AUC value increases by 1.42% compared to existing models.

Long-term Series Forecasting Method Based on Multi-granularity Multi-scale Deep Spatio-TemporalModeling

CHEN Jiahao, XIE Liang, LIAO Sihao, WU Yuchen, XU Haijiao

Computer Science. 2025, 52 (2): 91-98. doi:10.11896/jsjkx.240400127

Abstract

PDF(2285KB) ( 163 )

References | Related Articles | Metrics

Time series modeling has been the focus of research in a number of fields,including finance and transportation,and spatio-temporal models have received a lot of attention from researchers because of their ability to capture the complex associations and trends in time-series data more comprehensively.In recent years,long-term series forecasting based on spatio-temporal modeling has achieved remarkable results,but the existing methods are limited by multi-granularity or multi-scale studies,which cannot fully mine the spatio-temporal information of the data.To overcome this problem,a multi-granularity multi-scale deep spatio-temporal model(MMDSTM) is proposed.The model first obtains seasonal,periodic and granularity sequences by decomposing the initial data.Then,the multi-scale isometric convolution generates scale predictions,while attention-based spatio-temporal feature layers generates multi-granularity predictions.Finally,the prediction results of multi-granularity and multi-scale predictions are merged by multi-level fusion.In experiments,MMDSTM's MSE metric decreases by 6.2%,21.5% and 1% compared to other methods on stock,traffic and battery datasets,and the introduction of multi-granularity and multi-scale significantly improves the time series forecasting performance.

Check-in Trajectory and User Linking Based on Natural Language Augmentation

WANG Tianyi, LIN Youfang, GONG Letian, CHEN Wei, GUO Shengnan, WAN Huaiyu

Computer Science. 2025, 52 (2): 99-106. doi:10.11896/jsjkx.240600031

Abstract

PDF(2422KB) ( 116 )

References | Related Articles | Metrics

With the rapid development of positioning technology and sensors,user movement trajectory data is becoming increa-singly abundant but scattered on different platforms.In order to fully utilize these data and accurately reflect users' real beha-vior,the study of trajectory user linking has become crucial.This task aims to accurately correlate user identities from massive check-in trajectory data.In recent years,researchers have tried to use methods such as recurrent neural networks and attention mechanisms to deeply mine trajectory data.However,current methods face two major challenges when processing user check-in sequences.First,the limited spatiotemporal features in the check-in data are insufficient to comprehensively model check-in point information from both subjective and objective perspectives.Second,the topic of the user check-in sequence will affect understan-ding and modeling check-in sequences.In response to these two challenges,this paper proposes a trajectory user linking model based on natural language augmentation named NLATUL,and designs a set of natural language templates and soft prompt tokens to describe the check-in sequence,and uses the language model to understand the subjective intention in the check-in points,integrating the user's spatiotemporal status,and providing a new perspective and representation that fully models the check-in points from both subjective and objective aspects.On this basis,this paper infer the topic of the check-in sequence through prompt learning,and performs bi-direction encoding on the trajectory represented by the modeled check-in points,so as to achieve an accurate understanding of the user's check-in sequence through the combination of the check-in sequence topic and the check-in sequence encoding,which can link the trajectory with the user more effectively.Verified on two check-in datasets,the experimental results show that proposed method can more accurately link check-in trajectories and their corresponding users.

Node Classification Algorithm Fusing High-order Group Structure Information

ZHENG Wenping, HAN Yiheng, LIU Meilin

Computer Science. 2025, 52 (2): 107-115. doi:10.11896/jsjkx.240600091

Abstract

PDF(2619KB) ( 129 )

References | Related Articles | Metrics

There are usually high-order group structures with specific local connection patterns in local neighborhood of nodes,which can more accurately describe the topological characteristics of nodes and help to understand the structural characteristics of the network and the interaction patterns between nodes.The structural similarity between nodes can be computed using high-order group structural features within the local neighbors of a node,and a node classification algorithm is proposed based on fusing high-order group structure information(NHGS).Weisfeiler-Lehman(WL) algorithm is used to iteratively aggregate the label information of the k-tuple in a node's neighborhood to update its k-tuple label.The number of occurrences of nodes in different k-tuple labels constitutes the feature vector of nodes,and the cosine similarity between feature vectors is used to represent the structural similarity between nodes.Combined with the attribute information of the nodes,the node embedding is obtained through the autoencoder neural network,and then the nodes in the network are classified.NHGS combines the k-tuple node group structure information with the attribute information of a node to obtain the node representation containing the high-order structure information.Experiments on real attribute networks show that the proposed method can effectively calculate the structural similarity between nodes,and improve the performance of node classification tasks.

Multi-source-free Domain Adaptation Based on Source Model Contribution Quantization

TIAN Qing, LIU Xiang, WANG Bin, YU Jiangsen, SHEN Jiashuo

Computer Science. 2025, 52 (2): 116-124. doi:10.11896/jsjkx.240600004

Abstract

PDF(2476KB) ( 108 )

References | Related Articles | Metrics

As a new research direction in the field of machine learning,multi-source-free domain adaptation aims to transfer knowledge from multiple source domain models to the target domain,so as to achieve accurate prediction of target domain samples.Essentially,the key to solving multi-source-free domain adaptation lies in how to quantify the contribution of multiple source models to the target domain and utilize the diverse knowledge in the source models to adapt to the target domain.To address these issues,this paper proposes a multi-source-free domain adaptation method based on source model contribution quantization(SMCQ).Specifically,a source model transferability perception is proposed to quantify the transferability contribution of the source model,enabling the effective allocation of adaptive weights for target domain models.Additionally,an information maximization method is introduced to reduce cross-domain distributional discrepancies and mitigate model degradation.Subsequently,a credible partition global alignment approach is proposed to divide high-confidence and low-confidence samples to cope with the noisy environment caused by domain differences,effectively reduce the risk of incorrect label assignments.In addition,a sample local consistency loss is also introduced to mitigate the impact of pseudo-label noise on clustering errors of low-confidence samples.Finally,experiments conducted on multiple datasets validate the effectiveness of the proposed method.

Study on Integrated Model of Securities Illegal Margin Trading Accounts Identification Based on Trading Behavior Characteristics

ZUO Xuhong, WANG Yongquan, QIU Geping

Computer Science. 2025, 52 (2): 125-133. doi:10.11896/jsjkx.241000110

Abstract

PDF(2997KB) ( 122 )

References | Related Articles | Metrics

In response to issues with securities illegal margin trading accounts,such as the scarcity of public data,the unscientific feature selection,the limited research on identification methods that are lack of precision,an integrated model for the identification of securities illegal margin trading accounts based on trading behavior characteristics(CFS-RF-BP) is proposed.This integrated model consists of four main steps:dataset construction,feature extraction,feature selection,and account identification.According to the characteristics of securities account trading data,a simulated dataset is automatically generated,and the relevant features of securities illegal margin trading accounts are detailed labeled and calculated.Using a heterogeneous feature selection model,combined with methods such as logarithm and normalization,the importance of seventeen essential features primarily associated with securities illegal margin trading accounts is evaluated,and five of these features are selected to form a key feature set.Based on this,the BP neural network is employed to construct the classifier for the integrated model of securities margin trading account identification,optimizing key parameters such as weights and biases,thereby achieving automatic classification of securities illegal margin trading accounts.Simulation experiments indicate that the proposed CFS-RF-BP model has achieved excellent results in terms of feature selection,precision,recall rate,and processing efficiency of identification.

Distributed Two-stage Clustering Method Based on Node Sampling

ZHANG Manjing, HE Yulin, LI Xu, HUANG Zhexue

Computer Science. 2025, 52 (2): 134-144. doi:10.11896/jsjkx.240800040

Abstract

PDF(2740KB) ( 123 )

References | Related Articles | Metrics

To address the challenges of high computational resource consumption and low clustering efficiency in big data clustering scenarios,researchers have proposed an innovative distributed two-stage clustering method based on node sampling.This method first performs local clustering operations on the data at each local node,and then extracts representative data samples from each node based on the results of the local clustering.The selected sample data from each node is then transmitted to the central node.At the central node,further clustering analysis is conducted on the aggregated sample data,and the results of the sample clustering are transmitted back to the local nodes.Finally,each local node combines its own local clustering results with the sample clustering results to complete the final unification of clustering labels.Through this process,the two-stage clustering method has transformed the traditional centralized clustering algorithms into a more scalable,distributed model,and ensures the consistency of the clustering result to the global dataset.Theoretical analysis and experimental results both indicate that compared with the conventional full-data centralized clustering techniques,the two-stage clustering method offers a framework which effectively integrates the efficiency of parallel processing and the accuracy of integrated analysis.Without sacrificing the accuracy of clustering,it significantly improves clustering efficiency and reduces time costs,which provided a feasible and robust distributed solution tailored for the complexities inherent in big data clustering tasks.

Multi-view Clustering Based on Cross-structural Feature Selection and Graph Cycle AdaptiveLearning

XIN Yongjie, CAI Jianghui, HE Yanting, SU Meihong, SHI Chenhui, YANG Haifeng

Computer Science. 2025, 52 (2): 145-157. doi:10.11896/jsjkx.231100173

Abstract

PDF(3545KB) ( 128 )

References | Related Articles | Metrics

Most of the existing graph adaptive learning methods rely on high-dimensional raw data and inevitably have phenomena such as noise or missing information in the data,resulting in the inability to accurately select the important feature information in the high-dimensional data,in addition to ignoring the structural relevance of the multi-view representations in the feature selection process.To tackle the above problems,a multi-view clustering method(MLFS-GCA) based on cross-structural feature selection and graph cycle adaptive learning is proposed.First,a cross-structural feature selection framework is designed.By jointly learning the spatial structure characteristics of multi-view representations and the consistency of the clustering structure,the high-dimensional data is projected into a low-dimensional linear subspace,and the low-dimensional feature representation is learned with the assistance of view-specific basis matrix and consistent clustering structures.Second,a graph cycle adaptive learning module is proposed.The k nearest neighbors in the projection space are selected by the k-nearest neighbor(k-NN) method,and the similar structures are optimized cyclically in concert with matrix low-rank learning.Eventually,the shared sparse similarity matrix for clustering task is learned.The superiority of graph cycle adaptive learning in multi-view clustering is demonstrated through extensive experiments on various real multi-view datasets.

Game-theoretic Rough Group Consensus Decision-making Model Based on Individual-Whole SpanAdjustments and Its Applications

HOU Hanzhong, ZHANG Chao, LI Deyu

Computer Science. 2025, 52 (2): 158-164. doi:10.11896/jsjkx.240600044

Abstract

PDF(2442KB) ( 107 )

References | Related Articles | Metrics

Group consensus decision-making refers to the process in which a group of individuals adjust their opinions through collective negotiation to ensure that the problem is solved on the premise of reaching consensus.Exploring the group consensus model through the example of air quality assessments,this study first uses intuitionistic fuzzy numbers to represent individual evaluations and proposes a novel mapping function to convert real numbers into intuitionistic fuzzy numbers.Next,a method to adjust the relative span between individual and overall evaluations is proposed to achieve consensus,which helps quickly identify and adjust the differences between individual and overall evaluations.Then,based on the achieved consensus,a game-theoretic rough set model is employed to determine the threshold by balancing accuracy and generality.This approach improves performance by reducing the size of the boundary region,thereby increasing the accuracy of the decision results.Finally,the feasibility and effectiveness of the proposed method are validated through an air quality evaluation example.In conclusion,the proposed model not only enriches the related theoretical framework and effectively reduces the risk of group consensus decision-making,but also provides a feasible path for solving complex decision-making problems.

Attribute Reduction Algorithm Based on Fuzzy Neighborhood Relative Decision Entropy

XU Jiucheng, ZHANG Shan, BAI Qing, MA Miaoxian

Computer Science. 2025, 52 (2): 165-172. doi:10.11896/jsjkx.231100202

Abstract

PDF(2435KB) ( 119 )

References | Related Articles | Metrics

Aiming at the problem that fuzzy neighborhood rough set is sensitive to data distribution and cannot effectively evaluate the classification uncertainty of datasets with large density differences,this paper proposes an attribute reduction algorithm based on fuzzy neighborhood relative decision entropy.Firstly,the classification uncertainty of the sample is defined by using the relative distance,thus remodeling the fuzzy neighborhood similarity relationship.Combined with the variable precision fuzzy neighborhood rough approximation,the possibility of the sample being classified into the wrong category is reduced.Secondly,the information entropy is augmented with the fuzzy neighborhood credibility and coverage under the information view,and this is integrated with the fuzzy neighborhood relative dependence constructed based on the algebraic view to introduce the fuzzy neighborhood relative decision entropy.Finally,an attribute reduction algorithm based on the fuzzy neighborhood relative decision entropy is designed to evaluate the importance of attributes from both the information and algebraic viewpoints.Comparative experiments with six existing attribute reduction algorithms on eight public datasets show that the proposed algorithm can effectively measure the uncertainty of samples under different data distributions and improve the classification performance of data.

Illumination-aware Infrared/Visible Fusion for Object Detection

CHENG Qinghua, JIAN Haifang, ZHENG Shuaikang, GUO Huimin, LI Yuehao

Computer Science. 2025, 52 (2): 173-182. doi:10.11896/jsjkx.240300068

Abstract

PDF(3672KB) ( 132 )

References | Related Articles | Metrics

The method based on infrared/visible light fusion can effectively improve the effect of target detection in open scena-rios such as road traffic and security monitoring.The existing methods rarely design feature interaction mechanisms for infrared/visible light differences,which limits the accuracy and robustness of detection.Therefore,this paper designs an infrared/visible image fusion network based on dual-stream structure,which fully considers the differences between different modal images,and realizes accurate target recognition in the open environment by extracting and fusing the multi-level feature information of diffe-rent modal images.In order to solve the problem that the quality of visible image is easily affected by the change of ambient illumination,a lightweight illumination-aware module is designed,and the weight of infrared/visible fusion is dynamically adjusted through the weight allocation function,so as to improve the adaptability and accuracy of the fusion algorithm.At the same time,a parameter-free 3D attention module is designed to automatically identify the channel and spatial importance of the extracted features of the network,and different fusion weights are assigned according to the importance of different modes,which can improve the effect of network fusion without increasing the number of parameters of the network.In addition,this paper constructs a set of near-infrared/visible light datasets(NRS),which provides more source data for target recognition tasks in open scenes.Finally,a series of tests are carried out on the self-constructed dataset NRS and the public dataset M³FD,and the results show that the detection accuracy of the proposed method reaches 93.5% and 92.2%(mAP_.50) respectively,which can provide support for accurate target detection and recognition in open scenes.

Remote Sensing Change Detection Based on Contextual Fine-grained Information Restoration

DU Qiangang, PENG Bo, CHI Mingmin

Computer Science. 2025, 52 (2): 183-190. doi:10.11896/jsjkx.240400131

Abstract

PDF(3911KB) ( 120 )

References | Related Articles | Metrics

Remote sensing change detection plays a crucial role in both military and civilian fields.However,there is a large amount of pseudo-change noise due to the huge spatial and temporal gaps in data acquisition of change detection image pairs.Exis-ting change detection methods are based on learning object features from a dual-stream twin network,followed by pseudo-change noise removal via a series of proprietary networks.However,this mutually independent denoising approach lacks the ability to capture the interdependencies between image pairs,and often results in the loss of a large amount of fine-grained information due to excessive focus on the denoising design.The CFIR proposed in this paper mitigates the problem of fine-grained information loss by exploiting the data structure features of the image pairs to augment the model's ability to learn the contextual dependencies and to compensate for the lost fine-grained information.In addition,it employs a gating mechanism that eliminates pseudo-change noise in the change detection task and guides the network to extract relevant change features,mitigating the impact of extreme data imbalance in change detection on the model's ability to learn real changes.CFIR has demonstrated competitive performance in several change detection benchmarks.Compared with the state-of-the-art method,it improves F1 by 0.21% and IoU by 0.38% on the LEVIR-CD dataset,and improves F1 by 0.99% and IoU by 2.43% on the WHU-CD dataset.

Face Detection Algorithm Based on Multi-task Joint Learning in Weak Light Scenes

ZHANG Xia, SU Zhaohui, CHEN Lu

Computer Science. 2025, 52 (2): 191-201. doi:10.11896/jsjkx.231100166

Abstract

PDF(4291KB) ( 132 )

References | Related Articles | Metrics

In weak light scenes,face detection refers to the use of image processing techniques to detect faces.Currently,most face detection algorithms in weak light environments typically enhance the weak light images before performing face detection,neglecting the feature correlation between face detection and image enhancement,thereby affecting the generalization ability of the model.Inspired by the EnlightenGAN algorithm,this paper proposes a Multitask joint learning algorithm for face detection in weak light environments.First,it integrates the input layer shared representation of face detection and image enhancement.Se-cond,it combines the face attention network with EnlightenGAN,adding a local discriminator for face region determination based on the global-local discriminator.Finally,it introduces illumination weight parameters on the basis of self-regularized attention maps,adjusting them to optimize the accuracy of face detection.Experimental results on the DARK FACE dataset demonstrate that,compared with existing algorithms,the proposed algorithm achieves a 1.92% improvement in face detection accuracy,while also effectively enhances the visual quality of images captured under weak light conditions.

Unsupervised Multi-class Anomaly Detection Based on Prototype Reverse Distillation

HE Liren, PENG Bo, CHI Mingmin

Computer Science. 2025, 52 (2): 202-211. doi:10.11896/jsjkx.240400048

Abstract

PDF(3472KB) ( 109 )

References | Related Articles | Metrics

Unsupervised anomaly detection is widely used in industrial quality inspection and other domains due to its requirement of only normal samples for training.Existing single-class anomaly detection methods exhibit a significant performance decrease when directly applied to multi-class anomaly detection.Among them,knowledge distillation-based anomaly detection methods distill the feature knowledge of pre-trained teacher models on normal samples into student models.However,they have the problem that they can't guarantee thestudent models learn only normal sample knowledge in multi-class anomaly detection.This paper proposes an unsupervised multi-class anomaly detection method,PRD(Prototype based reverse distillation),based on a reverse knowledge distillation framework.It utilizes the Multi-class Normal Prototype module and Sparse Prototype Recall training stra-tegy to learn prototypes of multiple-class normal sample features from the teacher model.These prototypes are then used to filter the input features of the student model,ensuring that the student model only learns the feature knowledge of normal samples from the teacher model.PRD surpasses existing state-of-the-art methods on various industrial anomaly detection datasets.Qualitative,quantitative,and ablation experiments validate the effectiveness of the PRD framework and its internal modules.

Improved Mesh Optimization-based Image Stitching Algorithm of Large Field Binocular Vision

DONG Hui, ZHANG Yuansong, LIN Wenjie, WU Xiang, GUO Fanghong, ZHANG Dan

Computer Science. 2025, 52 (2): 212-221. doi:10.11896/jsjkx.231200068

Abstract

PDF(8555KB) ( 116 )

References | Related Articles | Metrics

To solve the problems of poor stitch quality caused by color inconsistency,distortion and artifacts in binocular image stitch,a large field binocular image stitch algorithm based on improved mesh optimization is proposed.Firstly,an improved mesh optimization method based on translation transformation and global alignment constraints is designed.By minimizing an objective function consisting of point-line alignment term,global alignment term and salient line preservation term,the optimal mesh vertex set is obtained,which achieves image registration while preserving the original shape and structure information.Secondly,an image fusion algorithm based on improved histogram matching and seam search is designed.The difference of brightness and hue in the overlapping area and the color cast phenomenon under large disparity are eliminated by improved histogram matching,and the seam search method based on human perception is used to obtain the weighted fusion after the suture line,which can effectively avoid the artifacts caused by using weighted fusion in the unaligned feature sparse area.Finally,the proposed algorithm is compared with four algorithms such as SPW,LPC,PSC and NOVATEK in 10 large-field scenes.The average registration error of the proposed algorithm is reduced by 28.1% and the average distortion error is reduced by 99.5% compared with the suboptimal algorithm.The results show that the proposed algorithm can not only effectively eliminate the hue difference between binocular images,but also suppress the projection distortion in the non-overlapping area of the target image and remove the artifacts in the overlapping area,which has obvious advantages.

Case Element Association with Evidence Extraction for Adjudication Assistance

LIU Yanlun, XIAO Zheng, NIE Zhenyu, LE Yuquan, LI Kenli

Computer Science. 2025, 52 (2): 222-230. doi:10.11896/jsjkx.240600081

Abstract

PDF(2371KB) ( 114 )

References | Related Articles | Metrics

Researchers in the past have devoted themselves to finding similar cases through the method of case matching.But the case-matching method depends on the text similarity.Similarity of texts is not equal to similarity of cases.Moreover,case ma-tching lacks interpretability.To address the shortcomings of case matching,we define a new problem,case element association with evidence extraction,which aims to predict the association results by elements rather than text similarity,and extracts factual details as evidence to explain the association result.This new problem is more in line with the actual needs of legal practitioners.In order to make the proposed model perform better on this new problem,contrastive learning is introduced to solve the problem of over-dependence on direct expressions of elements,which makes the attention weights evenly distributed on different expressions of same elements,thereby improving the effect of our model.We perform experiments on public and self-constructed datasets.Experiment results show that compared with text matching models,the proposed model improves the accuracy and precision by about 20%,and improves the recall and F1 by about 30%.

Role-aware Speaker Diarization in Autism Interview Scenarios

WANG Kangyue, CHENG Ming, XIE Yixiang, ZOU Xiaobing, LI Ming

Computer Science. 2025, 52 (2): 231-241. doi:10.11896/jsjkx.240100059

Abstract

PDF(1758KB) ( 108 )

References | Related Articles | Metrics

Speaker diarization technology plays a pivotal role in the field of intelligent speech transcription,with its core task being the segmentation and clustering of multi-speaker audio based on speaker identities,thereby facilitating better organization of audio content and transcribed text.In the scenarios of medical interview,speaker diarization technology serves as a prerequisite for subsequent automated assessment.Role information is naturally present in the field of medical interactive dialogue,taking autism as an example,the typical situation includes three well-defined roles:doctor,parent,and child undergoing diagnosis.However, in actual conversation,the correspondence between the role and the speaker may not always be one-to-one.For instance,during autism diagnosis,each conversation may involve only one child,while the number of doctors or parents may vary.We believe that the role information and the speaker information embedded in each speech segment can effectively complement each other,thereby reducing the diarization error rate.In this study,we propose a method integrating role information into the sequence-to-sequence target speaker voice activity detection(Seq2Seq-TSVAD) framework,achieving a diarization error rate(DER) of 20.61% on the CPEP-3 dataset.This error rate is 9.8% lower compared to the Seq2Seq-TSVAD baseline method and 19.3% lower compared to the conventional modular speaker diarization method,underscoring the significant effect of role information in enhancing speaker diarization performance in autism interview scenarios.

Joint Scheduling Algorithm of Battery Charging Power and User Allocation for Time-varyingElectricity Prices

WAN Desheng, CHEN Hao, CHENG Wenhui, GAO Yunlong

Computer Science. 2025, 52 (2): 242-252. doi:10.11896/jsjkx.240200018

Abstract

PDF(3387KB) ( 111 )

References | Related Articles | Metrics

With the global proliferation of electric motorcycles,battery-swapping stations have gained widespread attention due to their advantages,such as swift replenishment,convenience,and safety.Nonetheless,a predominant challenge most stations face is the elevated cost of charging.It is derived from the charging strategies for battery and user behavior during battery swaps.Leve-raging the adjustable charging power of batteries at swapping stations and the idle time of numerous batteries during the charging process is anticipated to reduce charging costs judiciously.Hence,this paper concentrates on the batteries and users within the battery swapping industry,specifically delving into the intricacies of adjusting the charging power of batteries and the allocation of users in battery swapping stations.The primary challenges encompass the dual influence of time-varying electricity prices and user allocation on battery charging power adjustment strategies.Furthermore,the battery and user allocation schemes must cater to users' battery swapping requirements across different time periods while also consider the charging costs incurred by individual batteries under varying charging power adjustment strategies.Consequently,adjusting battery charging power and user allocation are interlinked and form a mutually dependent problem.To this end,we propose a comprehensive scheduling algorithm for adjusting battery charging power and user allocation,taking into account the time-varying electricity prices.Firstly,a greedy strategy is initially applied to formulate a charging power adjustment plan for each battery.Then,considering the charging costs and the energy swapped out by individual batteries,an algorithm based on genetic characteristics is utilized to match batteries with optimal users,thereby minimizing the overall charging costs of the battery swapping station.Finally,the proposed method undergoes a thorough evaluation using a large-scale dataset from 44 battery swapping stations and 7 334 batteries in Chengdu,spanning two years.Experimental results demonstrate that,on average,the total charging costs of the proposed algorithm are reduced by 20.8% compared to the three baselines.

Dependency Parsing for Chinese Electronic Medical Record Enhanced by Dual-scale Collaboration of Large and Small Language Models

XU Siyao, ZENG Jianjun, ZHANG Weiyan, YE Qi, ZHU Yan

Computer Science. 2025, 52 (2): 253-260. doi:10.11896/jsjkx.231200054

Abstract

PDF(2331KB) ( 178 )

References | Related Articles | Metrics

Dependency parsing is a crucial task in natural language processing,aiming to identify the syntactic dependencies between words in a sentence.However,existing research on dependency parsing for Chinese electronic medical records faces follo-wing problems:current general-purpose parsers are unable to accurately analyze the situation when there is a lack of components indicative of grammatical structure and a variety of positions of modifiers.To address these issues,this paper proposes a method based on a dual-scale collaborative enhancement of large and small language models for dependency parsing of Chinese electronic medical records.Specifically,we first analyze the linguistic features of Chinese electronic medical records,and propose component completion to indicate special grammatical structures in medical texts.Subsequently,we utilize a generic parser for dependency parsing,for the parsed syntactic graph,we employ the prior grammatical knowledge of a large language model to modify it automatically.In addition,since our approach focuses on narrowing the feature distribution gap between medical and generic texts,it is not constrained by the lack of annotated data in the medical domain.This study annotates 444 samples for dependency parsing of Chinese electronic medical records,which validates our method.Experimental results demonstrate the effectiveness of our approach in parsing Chinese electronic medical records,achieving LAS and UAS metrics of 92.42 and 94.60 in the scenario with little data.The proposed method also shows significant performance in various departments.

Two-stage Multi-factor Algorithm for Job Runtime Prediction Based on Usage Characteristics

SHANG Qiuyan, LI Yicong, WEN Ruilin, MA Yinping, OUYANG Rongbin, FAN Chun

Computer Science. 2025, 52 (2): 261-267. doi:10.11896/jsjkx.240200072

Abstract

PDF(2524KB) ( 174 )

References | Related Articles | Metrics

To address the chain impact of inaccurate user-set job runtime on the scheduling system of high-performance computing platforms,a versatile two-stage multi-factor(TSMF) algorithm for job runtime prediction is proposed.TSMF integrates intricate user behavior patterns and nuanced job contextual features to ensure accurate and reliable predictions.TSMF can seamlessly embed into the scheduling systems of most high-performance computing platforms,thereby enhancing their performance.The multi-angle simulation experiments on the dataset and real scheduling system of Peking University's high-performance computing clusters show that TSMF performs well in prediction accuracy and can achieve accurate prediction on most jobs.For example,in up to 60.8 % of jobs,the prediction error is as low as less than one minute.Furthermore,TSMF significantly enhances the sche-duling algorithms in practical scenarios,improving resource utilization and substantially reducing user waiting times.

Flow Cardinality Estimation Method Based on Distributed Sketch in SDN

WANG Yijie, GAO Guoju, SUN Yu'e, HUANG He

Computer Science. 2025, 52 (2): 268-278. doi:10.11896/jsjkx.240900059

Abstract

PDF(3364KB) ( 129 )

References | Related Articles | Metrics

In software-defined networks(SDN),statistical flow cardinality information plays a crucial role in various applications such as traffic engineering,traffic rerouting,and attack detection.Existing research primarily falls into two categories:measurement structures deployed on a single switch and collaborative measurements across multiple switches.However,neither of the two approaches achieves full network flow coverage,and collaborative measurements often employ independent measurements from each switch followed by merging,which can lead to duplicate counting.To address these issues,a flow cardinality estimation method based on distributed sketches is proposed.This method leverages the advantages of centralized control in the SDN control plane and collaboratively utilizes switches along the longest continuous common sub-path of each flow to construct a logical layer counting structure for that flow.Additionally,a mapping from the logical layer counting structure to the physical switch space for each flow is established,allowing participating switches to dynamically adjust the mapping intervals based on their own states and actual load conditions,thus achieving load balancing among all switches in the network.Taking the vHLL algorithm as an example,we implement a prototype of the distributed flow cardinality estimation method and conduct experimental evaluations using a real network traffic dataset on a four-layer Fat-Tree network topology.The experimental results demonstrate that the proposed method effectively achieves flow cardinality estimation across the entire network.In terms of accuracy,its average relative error(ARE) and average absolute error(AAE) values outperform the comparison experiments by up to 94.7% and 93.8%,respectively.Regarding load balancing,the method fully utilizes all switches in the network,achieving a normalized average packet load of 0.394,which is lower than that of the comparison methods,indicating its good performance.

Fully Distributed Event Driven Bipartite Consensus Algorithm Based on Reinforcement Learning

CAI Yuliang, LYU Chunhui, HE Qiang, YU Bo, CHEN Dongyue, WANG Youtong, WANG Qiang, LIU Yuxuan, ZHAO Jingjing

Computer Science. 2025, 52 (2): 279-290. doi:10.11896/jsjkx.240100133

Abstract

PDF(2869KB) ( 115 )

References | Related Articles | Metrics

Reinforcement learning(RL) methods and fully distributed event driven control strategies are used to study the bipartite consensus problem of multi-agent systems(MASs) with unknown system model information.Firstly,a hybrid event triggered mechanism based on state threshold and time threshold is proposed to reduce the communication frequency between intelligent agents.Secondly,an adaptive event triggered consensus control protocol is designed using locally sampled state information,resulting in the consensus error of all following agents eventually approaching zero.The effectiveness of the above event triggered mechanism is confirmed by excluding Zeno behavior within a limited time.Then,based on the RL method,a model free algorithm is proposed to obtain the feedback gain matrix,and an adaptive event triggered control strategy is constructed in the presence of unknown model information.Unlike existing related works,the RL-based event triggered adaptive control algorithm only relies on locally sampled state information and is independent of any model information or global network information.In addition,we extend the above results to the switching topology scenario,which is more challenging because the state estimation is updated in the following two situations:1)when the interaction graph switches or 2)when the event triggering mechanism is satisfied.Finally,the effectiveness of the adaptive event triggered control algorithm is verified through examples.

Task Scheduling in Heterogeneous Server Systems Based on Data Splitting and Energy-aware Strategies

YANG Chen, XIAO Jing, WANG Mi

Computer Science. 2025, 52 (2): 291-298. doi:10.11896/jsjkx.241000027

Abstract

PDF(2022KB) ( 104 )

References | Related Articles | Metrics

Heterogeneous server platforms provide powerful computing capabilities for large systems but also pose challenges in system complexity and energy consumption management.This study delves into the energy-aware scheduling problem based on data splitting for dependent tasks in heterogeneous server systems.First,the system environment,dependent tasks,and data transmission patterns are modeled,and the energy-aware scheduling problem is formulated as a constrained optimization problem aimed at minimizing the completion time of task scheduling.Subsequently,an energy-aware scheduling algorithm(DSEA)based on data splitting and task prioritization strategies is proposed.This algorithm seeks approximate optimal startup times and server allocation plans for each task by optimizing data splitting strategies,task priorities,and weight-based energy allocation.To validate the effectiveness of the proposed method,1 000 jobs of varying lengths are randomly selected from the Alibaba cluster dataset for simulation experiments.Experimental results demonstrate that the DSEA algorithm exhibits significant performance advantages over three existing algorithms in various application scenarios.

Adaptive Operator Parallel Partitioning Method for Heterogeneous Embedded Chips in AIoT

LIN Zheng, LIU Sicong, GUO Bin, DING Yasan, YU Zhiwen

Computer Science. 2025, 52 (2): 299-309. doi:10.11896/jsjkx.240900101

Abstract

PDF(2609KB) ( 123 )

References | Related Articles | Metrics

With the continuous improvement of people's quality of life and the rapid development of technology,mobile devices such as smartphones have achieved widespread popularity globally.Against this backdrop,the deployment and application of deep neural networks on mobile devices have become a research hotspot.Deep neural networks not only drive significant progress in the field of mobile applications,but also pose higher requirements for energy efficiency management of battery-powered mobile devices.Meanwhile,the rise of heterogeneous processors in today's mobile devices brings new challenges to energy efficiency optimization.The allocation of computing tasks among different processors to achieve parallel processing and acceleration of deep neural networks does not necessarily optimize energy consumption and may even increase it.To address this issue,this paper proposes an energy-efficient adaptive parallel computing scheduling system for deep neural networks.This system comprises a runtime energy consumption analyzer and an online operator partitioning executor,which can dynamically adjust operator allocation based on dynamic device conditions,ensuring optimized energy efficiency for computing on heterogeneous processors of mobile devices while maintaining high responsiveness.Experimental results demonstrate that compared to baseline methods,the system designed in this paper reduces average energy consumption and latency by 5.19% and 9.0% respectively,and the maximum energy consumption and latency are reduced by 18.35% and 21.6% on mobile device deep neural networks.

Task Scheduling Strategy Based on Improved A2C Algorithm for Cloud Data Center

XU Donghong, LI Bin, QI Yong

Computer Science. 2025, 52 (2): 310-322. doi:10.11896/jsjkx.240500111

Abstract

PDF(4268KB) ( 128 )

References | Related Articles | Metrics

The existing task scheduling algorithms based on deep reinforcement learning(DRL) in cloud data center have the following problems.High training cost caused by low effective experience utilization,learning oscillation caused by variable and high dimension of state space,and slow convergence speed caused by fixed step size of policy update.In order to solve the above pro-blems,this paper constructs a parallel task scheduling framework based on the cloud data center scenario,and studies the cloud task scheduling problem with the goal of delay,energy consumption and load balancing.Based on DRL algorithm A2C,this paper proposes a task scheduling algorithm for cloud data center based on adaptive state optimization and dynamic step size A2C(AODS-A2C).Firstly,the admission control and priority strategy are used to filter and sort the queue tasks to improve the utilization of effective experience.Secondly,the dynamic high-dimensional state is quickly optimized in an adaptive way to maintain a relatively stable state space and avoid the oscillation problem in the training process.Finally,JS(Jensen Shannon) divergence is used to measure the probability distribution difference between the old and new strategies,and the learning step size of Actor network and Critic network is dynamically adjusted according to this difference,so as to quickly adjust to the best value for the current learning state and improve the convergence speed of the algorithm.The simulation results show that the proposed AOS-A2C algorithm has the characteristics of fast convergence speed and high robustness.Compared with other comparison algorithms,the delay is reduced by 1.2% to 34.4%,and the energy consumption is reduced by 1.6% to 57.2%,and it can achieve good load ba-lancing.

Research Progress in Facial Presentation Attack Detection Methods Based on Deep Learning

SUN Rui, WANG Fei, FENG Huidong, ZHANG Xudong, GAO Jun

Computer Science. 2025, 52 (2): 323-335. doi:10.11896/jsjkx.240200015

Abstract

PDF(2173KB) ( 121 )

References | Related Articles | Metrics

With the widespread application of facial recognition technology in fields such as public security and financial payments,presentation attacks(PAs) pose a serious challenge to the security of facial recognition systems.Presentation attacks detection(PAD) technology aims to determine the authenticity of input faces and has important research significance for maintaining the security and robustness of recognition systems.Due to the continuous emergence of large-scale datasets in recent years,deep learning-based PAD methods have gradually become the mainstream in this field.This paper offers a survey of current face PAD techniques based on deep learning.Firstly,it provides an overview of the definition,implementation methods,and common types of attack for PAD.Secondly,from the perspectives of single modality and multimodality,a thorough study is performed on the development trends,technical principles,benefits,and drawbacks of deep learning methods in the field of PAD over the previous five years.Thirdly,the common datasets and their characteristics that are used in PAD research are presented,and the evaluation standards,protocols,and algorithm performance results are given.Finally,we summarize the main issues faced in current PAD research and look forward to future development trends.

Generation Method for Adversarial Networks Traffic Based on Universal Perturbations

DING Ruiyang, SUN Lei, DAI Leyu, ZANG Weifei, XU Bayi

Computer Science. 2025, 52 (2): 336-343. doi:10.11896/jsjkx.240300031

Abstract

PDF(2173KB) ( 111 )

References | Related Articles | Metrics

Artificial intelligence technology has shown great potential in the field of network traffic classification and has had a profound impact on the strategic landscape of cyberspace security.But some studies have also found that deep learning models have serious vulnerabilities,and adversarial samples derived from this vulnerability can significantly reduce the accuracy of model detection.At present,adversarial samples are widely and deeply studied in the field of image classification,and are still in the development stage in the field of network traffic classification.The existing adversa-rial network traffic technology is only effective for specific samples,and has high time cost and low practicality.Therefore,this paper proposes a method for generating adversarialnetwork traffic based on general perturbations.It uses the properties of spatial feature distribution to find a general perturbation vector,adds this perturbation to normal traffic to generate adversarial network traffic,and causes a high probability of detection errors in the network traffic classifier.Compared with existing methods,this paper conducts experimental tests on Moore dataset and ISCX2016 dataset.The results show that under the same conditions,this method is effective for generating adversarial network traffic attack classifiers for all samples on Moore dataset and ISCX2016 dataset,with a success rate of over 80%.It can effectively attack different classifiers,with model transferability effect.At the same time,the time cost is less than 1 ms,achieving rapid generation of adversarial network traffic with much higher efficiency than existing methods.

Augmenter:Event-level Intrusion Detection Based on Data Provenance Graph

SUN Hongbin, WANG Su, WANG Zhiliang, JIANG Zheyu, YANG Jiahai, ZHANG Hui

Computer Science. 2025, 52 (2): 344-352. doi:10.11896/jsjkx.240400029

Abstract

PDF(1991KB) ( 113 )

References | Related Articles | Metrics

In recent years,advanced persistent threat(APT) attacks have become increasingly prevalent.Data provenance graphs,which contain rich contextual information reflecting process execution,have shown potential for detecting APT attacks.Therefore,provenance-based intrusion detection systems(PIDS) have garnered attention.PIDS identify malicious behavior by capturing system logs to generate provenance graphs.PIDS encounter the following main challenges:efficiency,generality,and real-time capability,particularly in terms of efficiency.Current PIDS generate thousands of alerts for a single anomalous node or graph,lea-ding to a significant number of false positives,which inconveniences security personnel.This paper presents Augmenter,the first PIDS simultaneously addresses the three aforementioned challenges.Augmenter partitions processes into communities based on the information fields of nodes,effectively learning the behavior of different processes.Additionally,Augmenter introduces a time-window strategy for subgraph partitioning and employs an unsupervised feature extraction method based on graph mutual information maximization.The incremental feature extraction algorithm amplifies abnormal behavior and distinguishes it from normal behavior.Finally,Augmenter trains multiple clustering models based on process types to achieve event-level detection,allowing for more precise localization of attack behaviors.Augmenter is evaluated on the DARPA dataset,confirming its real-time performance by measuring the efficiency of the detection phase.In terms of detection efficiency,we compare the precision and recall rates with the state-of-the-art works,Kairos and ThreaTrace.Kairos achieves precision and recall rates of 0.17 and 0.80,while ThreaTrace achieves 0.29 and 0.76.In contrast,Augmenter achieves precision and recall rates of 0.83 and 0.97,demonstrating that Augmenter has significantly higher precision and detection performance.

Improvement of SSH Transport Layer Protocol Based on Chain of Trust

WANG Xingguo, SUN Yunxiao, WANG Bailing

Computer Science. 2025, 52 (2): 353-361. doi:10.11896/jsjkx.231200187

Abstract

PDF(2129KB) ( 96 )

References | Related Articles | Metrics

Host keys are identification of SSH servers.Users are required to check host key fingerprints to authenticate SSH servers.However,users often ignore the process of checking fingerprints when using SSH,making man-in-the-middle attacks based on host key replacement possible.In this regard,an improvement scheme of the SSH transport layer protocol is proposed based on the chain of trust.In the scheme,a chain of trust is established by signing the new host key with the old host key.The improved SSH protocol can solve the trust problem of new host keys without the need for users to check fingerprints,so as to achieve identity authentication of servers,which greatly reduces the risk of man-in-the-middle attacks.Finally,using ProVerify to analyze the improved protocol,verification results show that the improved protocol satisfies confidentiality and authentication,and can resist man-in-the-middle attacks.

Traffic Adversarial Example Defense Based on Feature Transfer

HE Yuankang, MA Hailong, HU Tao, JIANG Yiming, ZHANG Peng, LIANG Hao

Computer Science. 2025, 52 (2): 362-373. doi:10.11896/jsjkx.240300009

Abstract

PDF(11194KB) ( 110 )

References | Related Articles | Metrics

In the domain of traffic detection,the challenge of defending against adversarial examples is significant.Traditional adversarial example defense methods,which rely heavily on adversarial training,necessitate a vast quantity of adversarial examples for training purposes.However,a notable drawback of such approaches is the resultant decrease in the recognition accuracy of the original,unaltered data.This reduction in accuracy poses a substantial problem,as it compromises the effectiveness of the defense mechanism in recognizing legitimate traffic patterns.To address these challenges,a novel approach to traffic adversarial example defense has been proposed,leveraging the concept of feature transfer.This innovative method ingeniously combines two strategic defense philosophies:firstly,enhancing the robustness of the model against adversarial attacks,and secondly,obfuscating the space within which adversarial examples operate.The defense mechanism is architecturally composed of two integral modules:a lower-level defense module equipped with denoising capabilities,and a recognition module designed for the explicit purpose of identifying traffic patterns.The cornerstone of this approach is the employment of a stacked autoencoder as the foundational element of the lower-level defense module.This choice is pivotal,as the autoencoder excels in adversarial knowledge learning,thereby endowing the system with the capability to extract and understand adversarial features effectively.This is a critical step in ensuring that the defense mechanism can preemptively neutralize potential adversarial threats.Subsequently,the system embarks on a phase of functional adaptation,tailored specifically to the characteristics of network traffic.This phase involves the construction of adaptive functionalities based on the distinct features of traffic,followed by the training of the recognition module using non-adversarial traffic data.This strategic training empowers the recognition module with the ability to accurately identify legitimate traffic patterns,thereby significantly enhancing the overall efficacy of the defense mechanism.A key innovation of this method is the conceptual separation of defense and recognition functionalities.This separation not only reduces the operational costs asso-ciated with defense but also minimizes the adverse impact of adversarial training on the recognition accuracy of original data.As a result,the system achieves a rapid adaptation to evolving threats,significantly improving the model's defensive resilience.Empirical evidence supports the effectiveness of this approach,with the recognition accuracy for new adversarial examples experiencing a substantial increase to approximately 40%.This improvement marks a significant advancement in the field of traffic detection and adversarial example defense,offering a promising avenue for future research and development.

Explanation Robustness Adversarial Training Method Based on Local Gradient Smoothing

CHEN Zigang, PAN Ding, LENG Tao, ZHU Haihua, CHEN Long, ZHOU Yousheng

Computer Science. 2025, 52 (2): 374-379. doi:10.11896/jsjkx.240400210

Abstract

PDF(2414KB) ( 110 )

References | Related Articles | Metrics

While the interpretability of deep learning is developing,its security is also facing significant challenges.There is a risk that the interpretation results of the model on input data may be maliciously manipulated and attacked,which seriously affects the application scenarios of interpretability technology and hinders human exploration and cognition of the model.To address this issue,an interpretable robust adversarial training method using model gradients as similarity constraints is proposed.Firstly,adversarial training data is generated by sampling along the interpretation direction.Secondly,multiple similarity metrics between the interpretations of the sampled data are calculated by combining the gradient information of the samples during the training process,which is used to regularize the model and smooth its curvature.Finally,to verify the effectiveness of the proposed interpretable robust adversarial training method,it is validated on multiple datasets and interpretation methods.The experimental results show that the proposed method has a significant effect on defending against adversarial interpretation samples.

Low-power Bluetooth Spoofing Attack Detection Technology Based on RFFAD_DeepSVDD

YAN Tingju, CAO Yan, WANG Yijing

Computer Science. 2025, 52 (2): 380-387. doi:10.11896/jsjkx.231200168

Abstract

PDF(2105KB) ( 101 )

References | Related Articles | Metrics

Aiming at the problem of low accuracy rate of existing low-power Bluetooth spoofing attack detection techniques,a BLE spoofing attack detection technique based on anomalous fingerprints is proposed,which takes the attacker's RF fingerprints as anomalous data and models spoofing attack detection as an anomalous detection problem.This paper designs an anomalous fingerprint detection model—RFFAD_DeepSVDD,based on deep support vector data description(DeepSVDD),which uses the residual unit to construct a network model,effectively alleviating the problem of insufficient nonlinear feature extraction of machine learning anomaly detection algorithm.Pre-trained auto-encoder is used to obtain the optimal initialization parameter,which greatly enhances the model's boundary decision-making ability.In the anomaly detection experiments,the accuracy of the model reaches 95.47%,an average improvement of 8.92% compared with the machine learning-based anomaly detection model;in the spoofing attack detection experiments,compared with the existing spoofing attack detection techniques in the attack node movement and stationary state,the proposed method shows better performance,and can accurately detect and identify man-in-the-middle attack,impersonation attack,and reconnection spoofing attackthree kinds of spoofing attacks.