Computer Science
Started in January,1974（Monthly）
Supervised and Sponsored by Chongqing Southwest Information Co., Ltd.
ISSN 1002-137X
CN 50-1075/TP
CODEN JKIEBK
Authors
Reviewers
Editors
Current Issue
Volume 52 Issue 5, 15 May 2025
  
High Performance Computing
AI+HPC:An Overview of Supercomputing System Software and Application Technology Development Driven by “AI+”
TAN Zhengyuan, ZHONG Jiaqing, CHEN Juan
Computer Science. 2025, 52 (5): 1-10.  doi:10.11896/jsjkx.241100177
Abstract PDF(2035KB) ( 199 )   
References | Related Articles | Metrics
Artificial Intelligence(AI) and High Performance Computing(HPC) are two essential technologies in computer science.With the rapid development of computer science and technology,there has been a gradual trend of convergence and deve-lopment of AI and HPC.On the one hand,new challenges in high-performance computing systems require AI-powered solutions(AI for HPC).On the other hand,breakthroughs in artificial intelligence demand the support of high-performance computing(HPC for AI).Consequently,the convergence of AI and HPC strikes the development of core technologies in their respective fields.In this paper,we systematically review the respective technological development in the fields of AI and HPC in the past decade,focusing on three aspects:1)the role of AI technology in HPC hardware architecture,operating system resource management,compilation optimization,and software development,etc;2)the support of HPC for AI in terms of system hardware solutions and software applications;3)prospects and challenges for the future development of AI and HPC convergence.
Performance Evaluation and Optimization of Operating System for Domestic Supercomputer
GAO Yiqin, LUO Zhiyu, WANG Yichao, LIN Xinhua
Computer Science. 2025, 52 (5): 11-24.  doi:10.11896/jsjkx.240500103
Abstract PDF(5755KB) ( 148 )   
References | Related Articles | Metrics
Supercomputers play a crucial role in supporting scientific computing applications.During these five years,our country is developing post-exascale domestic supercomputers.As one of the core components of supercomputers,the operating system's overhead will impact the performance of the supercomputer system.Therefore,the evaluation of the OS is one of the important subjects in supercomputer research.Among existing domestic OSs,openEuler offers high performance and compatibility on systems equipped with Kunpeng processors.However,openEuler has not been extensively applied to supercomputers.Therefore,it is necessary to evaluate its performance on supercomputers,and optimize the existing performance bottlenecks.Our work can be divided into two parts.1)We evaluate the compatibility of openEuler and its performance when running HPC applications.CentOS is used as a reference for comparison.The evaluation results show that when running non-communication-intensive applications,the performance of openEuler is comparable to CentOS.However,when using OpenMPI for collective communication operations such as Allreduce,the performance on openEuler decreases by up to 76.83%.Additionally,under thousand-core scale,the parallel efficiency of communication-intensive applications on openEuler decreases by up to 23.01%.2)Based on the performance issues with MPI collective communication identified during the evaluation process,we propose a performance modeling and optimization method.This method relies on the Hockney model of point-to-point communication to model the performance of various collective communication algorithm implementations.It predicts communication time under different numbers of processes and message sizes,enabling the selection of suitable collective communication algorithm implementations.Utilizing the MCA interface of OpenMPI,this method allows for dynamic adjustment of algorithm implementations at runtime.After optimization,the perfor-mance of HPC applications on openEuler has been significantly improved,with a maximum reduction in running time of 26%.
Accelerating Batched Matrix Multiplication for Variable Small Sizes Based on TVM andApplications
DAI Hanwen, CHEN Changbo
Computer Science. 2025, 52 (5): 25-40.  doi:10.11896/jsjkx.240500052
Abstract PDF(4815KB) ( 127 )   
References | Related Articles | Metrics
In many practical applications,efficient computation of a large amount of small matrix products across different dimensions is required.For instance,in graph classification tasks based on graph neural networks,multiple adjacency matrices need to be multiplied with node feature matrices.To address the issue of existing methods being unable to efficiently handle batched matrix multiplication for variable and small sizes across different hardware platforms,this paper introduces a cross-platform efficient algorithm,BVSM,based on the deep learning compiler TVM.BVSM enhances the capability of TVM to efficiently perform batched matrix multiplication for small and variable sizes by utilizing customized optimization templates for small matrices,employing tensorization for batching,and applying grouped padding techniques.Experiments on real datasets of graph classification demonstrate that,on the CPU,BVSM achieves on average over two times speedup compared to auto-scheduled and auto-tuned TVM(AnsorTVM),reaching 95% of the average performance of,and achieving up to 1.27 times compared to,the method of batched matrix multiplication for variable sizes of Intel MKL.On the GPU,BVSM achieves on average over 62.05 times speedup compared to AnsorTVM,over 28.82 times speedup compared to cuBLAS,and over 6.59 times speedup compared to the method of batched matrix multiplication for variable sizes of MAGMA.
Metrics and Tools for Evaluating the Deviation in Parallel Timing
LIAO Qiucheng, ZHOU Yang, LIN Xinhua
Computer Science. 2025, 52 (5): 41-49.  doi:10.11896/jsjkx.241200053
Abstract PDF(3624KB) ( 91 )   
References | Related Articles | Metrics
In parallel computing,instrumenting specific code segments is commonly used for performance evaluation on multicore processors.However,factors such as timing methods,hardware configurations,and runtime environments affect parallel timing accuracy,jeopardizing stability and reproducibility of performance measurements.As the core number of multicore processors grows,accurate parallel timing has grown more challenging.Two key problems remain:1)current method cannot quantitatively compare the accuracy of different timing methods;2)the root cause of parallel timing variability is not fully understood.This paper proposes metrics for evaluating the deviation in measurements and presents ParTES,a tool which emulates realistic cache conditions and timing intervals on X86 and Armv8 CPUs,allowing quantitative evaluation of timing variability across different timing methods.This study performed microsecond-level and millisecond-level analyses of parallel timing deviations on Kunpeng,Phytium,and Hygon processors.The results show that the performance of timing methods,cache status,nearby instructions,and server hardware configurations all influence accuracy is excellent.Among these CPUs,the most stable timing methods are PAPIon Kunpeng,POSIX's clock_gettime on Phytium,and the RDTSC instruction on Hygon.
Impact and Analysis of Optimizers on the Performance of Neural Network Force Fields
LI Enji, HU Siyu, TAN Guangming, JIA Weile
Computer Science. 2025, 52 (5): 50-57.  doi:10.11896/jsjkx.241100176
Abstract PDF(4086KB) ( 85 )   
References | Related Articles | Metrics
Molecular dynamics(MD) simulation is widely used in various fields,such as materials science and computational chemistry.In recent years,with the improvement in computational power,the development of neural network models,and the accumulation in first-principle data,neural network force field(NNFF) models have demonstrated high predictive accuracy.Curren-tly,there are multiple training algorithms available for NNFF models,and these models are undergoing rapid iteration.However,there remains a lack of guidance on NNFF models and their compatible optimizers.This paper selects three representative NNFF models and the three most commonly used optimization algorithms for these models,testing and evaluating them on four real-world datasets to analyze factors affecting their convergence.We have designed numerous experiments for a comprehensive evaluation,including the impact of model parameter size on the optimizer,the influence of model depth and width on convergence,and the relationship between model training time and the optimizer.Our work provides recommendations for optimizer algorithms specific to NNFF models.
Operator Fusion Optimization for Deep Learning Compiler TVM
GAO Wei, WANG Lei, LI Jianan, LI Shuailong, HAN Lin
Computer Science. 2025, 52 (5): 58-66.  doi:10.11896/jsjkx.240100018
Abstract PDF(2705KB) ( 67 )   
References | Related Articles | Metrics
Operator fusion technique is an optimization method employed by deep learning compilers to combine multiple operators into a single,larger operator.This approach effectively reduces computation costs and memory access requirements.In the operator fusion scheme of deep learning compiler TVM,operators are categorized based on their functional characteristics,fusion rules are devised,and a greedy algorithm is utilized for fusion.However,this fusion scheme has the following problems.Firstly,the fusion rules derived from functional feature classification may not be sufficiently generalizable,leading to missed opportunities for operator fusion and limited granularity.Secondly,the greedy algorithm fails to achieve optimal solutions for operator fusion.To address these issues,improvements are made in TVM by introducing an operator classification method based on input/output mapping types and designing a more comprehensive set of fusion rules that expand the granularity of operator fusion.Additionally,a search algorithm for finding suitable fusion schemes and a cost evaluation model based on dynamic programming are proposed to prune the search space and enable efficient identification of optimal solutions within reasonable time frames.To evaluate the effectiveness of this enhanced fusion scheme,experiments are conducted using popular deep learning models such as VGG-16,Efficient-B0,MobileNet-V1 and YOLO-V4 on both CPU and DCU platforms.The experimental results show that compared with the original fusion scheme of TVM,the fusion ratio of deep learning models can be improved.The average fusion ratio is increased by 27%,and the average inference delay rate is 1.75.
TS3:Energy-Efficiency-First Optimal Thread Number Search Algorithm Based on Specific Starting Point Classification
MA Zhaoyang, CHEN Juan, ZHOU Yichang, WU Xianyu, GAO Pengfei, RUAN Wenhao, ZHAN Haoming
Computer Science. 2025, 52 (5): 67-75.  doi:10.11896/jsjkx.241100175
Abstract PDF(4508KB) ( 47 )   
References | Related Articles | Metrics
Optimal thread number setting is one of the key factors affecting the performance and power consumption of multi-threaded programs.However,current algorithms for finding the optimal number of threads usually start the search from a single fixed point,which cause the problem of low precision and large search overhead.The distribution and location of the optimal number of threads are related to various,factors,including types of programs,optimization objectives(performance,power consumption,and EDP),parallel multi-threaded areas,and software-hardware configuration parameters.This paper focuses on the problem of searching for the optimal number of threads with an emphasis on energy efficiency and proposes an energy-efficiency-first optimal thread number search algorithm based on specific starting point classification(abbreviated as TS3 method).By designing a multi-threaded program classifier to optimize the setting of search starting points,and applying heuristic and binary search algorithms to enhance search efficiency,the method effectively improves the accuracy of the optimal number of threads search under energy efficiency priorities(optimal performance,optimal powerconsumption,optimal EDP) and reduces search costs.The effectiveness of the algorithm is experimentally validated using eight benchmarks on two x86 platforms and one ARM platform.Compared to the baseline,the TS3method achieves an average performance improvement of 0.29%(Platform A),0.17%(Platform B),and 10.77%(Platform C);average power consumption reduction of 2.35%(Platform A),1.87%(PlatformB),and 15.97%(Platform C);and average EDP reduction of 6.36%(Platform A),5.07%(Platform B),and 46.94%(Platform C).Across the three platforms,compared to current classical search methods,the TS3 method demonstrates an average performance improvement of 10.16%,an average reduction in power consumption of 13.45%,and an average reduction in EDP of 23.77%,the search overhead is reduced by 86.8%.
Research on Function Vectorization Technology Based on Directive Statements
LIU Lili, SHAN Zheng, LI Yingying, WU Wenhao, LIU Wenbo
Computer Science. 2025, 52 (5): 76-82.  doi:10.11896/jsjkx.231200174
Abstract PDF(1748KB) ( 70 )   
References | Related Articles | Metrics
With the continuous development of processor technology,SIMD(Single Instruction Multiple Data) vectorization has been widely applied in various fields.However,previous research has mainly focused on loops and basic blocks,while full-function vectorization can better exploit the advantages of SIMD instructions,thereby improving application performance.This paper proposes a guided statement-based function vectorization method.Firstly,a relatively simple guided statement is added to the loop involving function calls,which can vectorize the instructions involving function calls in the loop.Secondly,the vectorization of the called function is achieved by using full function vectorization to generate a vectorized full function instead of inlining it.Finally,the function call instructions in the loop are processed to generate vectorized function call instructions.We selected 10 benchmarks from ISPC benchmark tests and SIMD library benchmark tests to evaluate our method,and the experimental results show that compared to scalar,the average speedup achieved is 6.949 times.
Research on LLM Vector Dot Product Acceleration Based on RISC-V Matrix Instruction Set Extension
CHEN Xuhao, HU Sipeng, LIU Hongchao, LIU Boran, TANG Dan, ZHAO Di
Computer Science. 2025, 52 (5): 83-90.  doi:10.11896/jsjkx.241200074
Abstract PDF(1479KB) ( 71 )   
References | Related Articles | Metrics
Considering the high-performance and low-power requirements of edge AI,this paper designs a specialized instruction set processor for edge AI based on the RISC-V instruction set architecture,addressing practical issues in digital signal processing for edge devices.This design enhances the execution efficiency of edge AI and reduces its energy consumption with limited hardware overhead,meeting the demands for efficient large language model(LLM) inference computation in edge AI applications.For the characteristics of large language models,custom instructions were extended based on the RISC-V instruction set to perform vector dot product calculations,accelerating the computation of large language models on dedicated vector dot product acceleration hardware.Based on the open-source high-performance RISC-V processor core XiangShan Nanhu architecture,the vector dot product specialized instruction set processor Nanhu-vdot is implemented,which adds vector dot product calculation units and pipeline processing logic on top of the XiangShan Nanhu.The Nanhu-vdot underwent FPGA hardware testing achieves over four times of the speed of scalar methods in vector dot product computation.Using a hardware-software co-design approach for second-generation generative pre-trained Transformer(GPT-2) model inference,the speed improves by approximately 30% compared to pure software implementation with almost no additional consumption of hardware resources and power consumption.
Hardware-Software Co-design Fault-tolerant Strategies for Systolic Array Accelerators
WEI Xiaohui, GUAN Zeyu, WANG Chenyang, YUE Hengshan, WU Qi
Computer Science. 2025, 52 (5): 91-100.  doi:10.11896/jsjkx.240800055
Abstract PDF(5124KB) ( 63 )   
References | Related Articles | Metrics
In recent years,with the continuous improvement in model inference accuracy,convolutional neural networks(CNNs) have been widely applied in safety-critical fields.To meet the demands of CNNs for real-time,high-performance,and low-power computing,domain-specific CNN accelerators is proposed.Among these,systolic array architectures have been extensively used due to their simple structure and high parallelism.However,factors such as process variations and equipment aging make systolic arrays prone to Stuck-At faults(SAF),which can lead to catastrophic accidents.Therefore,fault-tolerant strategies for systolic arrays are critically important.Existing fault-tolerant strategies,however,suffer from high time and resource costs,as well as excessive modifications to network parameters.To achieve an efficient and low-overhead lightweight fault-tolerant strategy,this paper aims to exploit the inherent fault tolerance of CNNs by relaxing the handling of minor SAFs,thereby reducing overall fault-tolerance overhead.Additionally,by fully considering the computational characteristics of systolic arrays,this paper proposes two hardware-software co-design fault-tolerant strategies:row(column) swapping and weight splitting.These strategies effectively mitigate the impact of SAF on model inference accuracy.Experimental results show that,compared to traditional row(column) bypass and selective protection strategies,the proposed hardware-software co-design fault-tolerant strategies offer superior execution efficiency and model accuracy recovery.
Investigation on Load Balancing Strategies for Lattice Boltzmann Method with Local Grid Refinement
HUANG Chenxi, LI Jiahui, YAN Hui, ZHONG Ying, LU Yutong
Computer Science. 2025, 52 (5): 101-108.  doi:10.11896/jsjkx.241100169
Abstract PDF(2821KB) ( 53 )   
References | Related Articles | Metrics
The Lattice Boltzmann method(LBM) with local mesh refinement is widely used in large-scale unsteady computational fluid dynamics problems.Although the local mesh refinement method can effectively reduce the computational workload,it also brings serious load balancing issues.Especially in large-scale parallel computing,the choice of load balancing algorithms has a significant impact on the overall computational efficiency.This paper used a self-developed LBM program to compare the implementations of Palabos and to explore the static load balancing algorithm based on the Lattice Boltzmann method with local mesh refinement.This paper implementes various load balancing strategies based on global load and stratified load,uses greedy algorithms and space-filling curve algorithms,and proposes optimization of the load balancing of computing nodes.It uses the flow around a sphere and the DrivAer case as test cases for testing and comparison.Firstly,the results show the performance based on the stratified load strategy is better than that based on the global load strategy.At the same time,a comparative analysis reveals that the greedy algorithm demostrates superior scalability,whereas the space-filling curve algorithmexhibits higher efficiency when operating with a limited number of process.Finally,based on layered load,a hybrid algorithm is implemented by combining greedy algorithm and space filling curve algorithm,which achieves the best performance when the number of processes is larger.
Database & Big Data & Data Science
Review of Doctor Recommendation Methods and Applications for Consultation Platforms
WU Xingli, ZHANG Haoyue, LIAO Huchang
Computer Science. 2025, 52 (5): 109-121.  doi:10.11896/jsjkx.240600149
Abstract PDF(2216KB) ( 65 )   
References | Related Articles | Metrics
This paper aims to strengthen the use of personalized recommendation technologies in online medical settings,help patients choose resources for high-quality physicians,and address the information overload caused by the growing volume of online consultations.Firstly,bibliometrics summarizes popular research directions.On this basis,this paper sorts out the existing online doctor recommendation methods and classifies them into five categories based on doctor-patient matching:recommendation based on traditional recommendation algorithms,recommendation based on multi-attribute decision making,recommendation based on machine learning,hybrid recommendation,and others.In addition,we compare the application status,advantages and disadvantages,and the application scope of each category.Finally,we analyze the trend of online doctor recommendations and propose future research directions.Online doctor recommendation belongs to the intersection of research problems in the fields of computer science,management,and medicine.In contrast to traditional recommender systems,online doctor recommendation prioritizes precise matching between patients' conditions and doctors' specialties.Traditional recommendation algorithms are initially applied in doctor recommendation,but they are constrained by data sparsity and cold start.Recommendation based on multi-attribute decision making possess a solid theoretical foundation and can flexibly reflect patient preferences,yet they require a high level of interaction between the system and patients.Recommendation based on machine learning can alleviate the challenge of data sparsity and enable intelligent recommendation,though they necessitate large data support and often suffer from poor interpre-tability.Hybrid recommendation models,by integrating the strengths of various algorithms,have the potential to improve recommendation performance.However,the challenge lies in combining and balancing these algorithms.Other research directions such as recommendation grounded in optimization theory and graph models remain to be explored.In the future,it will be important to integrate multidisciplinary methodologies,conduct research on cross-platform,multi-source,and heterogeneous doctor-patient data mining,expression,and integration,and explore doctor recommendation modes based on patients' needs and preferences.
Study on EEG Emotion Recognition Method Based on Self-supervised Graph Network
ZHANG Jiaxiang, PAN Min, ZHANG Rui
Computer Science. 2025, 52 (5): 122-127.  doi:10.11896/jsjkx.240200039
Abstract PDF(2219KB) ( 72 )   
References | Related Articles | Metrics
EEG emotion recognition refers to the technology of identifying human emotional states by analyzing electroencephalogram (EEG) signals,which has wide application prospects in some fields such as medical health,and human-computer interaction.Currently,EEG-based emotion recognition frequently relies on machine learning or deep learning techniques to thoroughly train labeled EEG data and differentiate various emotional states.However,such methods require a lot of data annotation,which is time-consuming and labor-intensive.Meanwhile,research shows that the spatial structure information of EEG signals can reflect the interaction of brain areas related to different emotional states,which can help identify emotional characteristics.To this end,this paper proposes an EEG emotion recognition method based on self-supervised graph network.First,the meiosis method is used to preprocess the EEG signal.Then,a graph convolutional network is used to extract spatial structure information from EEG signals,and the network is trained through self-supervised tasks.Finally,the feasibility and effectiveness of the proposed method have been validated through numerical experiments using the public datasets SEED and SEED-IV.Numerical results show that the accuracy of emotion recognition is 95.16% and 80.23%,which is superior to current methods.
Point-of-interest Recommendation Based on Geospatial-TemporalCorrelations and Social Influence
JIN Hong, CHEN Like, YOU Lan, LYU Shunying, ZHOU Kaicheng, XIAO Kui
Computer Science. 2025, 52 (5): 128-138.  doi:10.11896/jsjkx.240200099
Abstract PDF(2388KB) ( 65 )   
References | Related Articles | Metrics
With the popularity of location-based social networks,personalized POI recommendation has become an important task.However,existing research inadequately considers the intrinsic relationships when modeling and integrating contextual information.Among these,geographical and temporal contexts often interact with and influence each other.Moreover,when mode-ling user social relationships,current approaches primarily measure the direct similarity between POI subsets visited by different users to express the similarity of their social behaviors.These approaches fails to effectively alleviate the impact of data sparsity on measuring the similarity of POI subsets visited by different users.Therefore,by reasonably reconstructing the contextual information model and effectively integrating it into user preference model,a POI recommendation method based on geospatial-temporal correlations and social influence is proposed.This method leverages the spatial aggregation phenomenon of users' primary geographic activity centers under different temporal states.It employs a time-constrained approach to assess the geographical correlations between POIs,thereby modeling the impact of POI geographical information on user check-ins.Furthermore,when mo-deling user social relationships,it is assumed that users with more shared check-ins at POIs or greater overlap in the categories of POIs they visit exhibit higher similarity in their social behaviors.By incorporating POI category information,the effectiveness and utility of social relationship modeling are enhanced.Finally,the proposed geospatial-temporal correlation model and user social relationship model are integrated into a weighted matrix factorization framework to perform personalized POI recommendations for users.Extensive experiments demonstrate that the proposed method achieves superior POI recommendation performance,highlighting the advantages of the proposed models in contextual modeling and overcoming data sparsity.
Study on Graph Collaborative Filtering Model Based on FeatureNet Contrastive Learning
WU Pengyuan, FANG Wei
Computer Science. 2025, 52 (5): 139-148.  doi:10.11896/jsjkx.240200078
Abstract PDF(3331KB) ( 60 )   
References | Related Articles | Metrics
Graph-based collaborative filtering recommendation techniques have gained significant attention for their ability to efficiently process large-scale interaction data.However,the effectiveness of these techniques is limited by the sparsity of data in real-world scenarios.Recent research has started to apply contrastive learning to graph collaborative filtering to enhance its performance.Nonetheless,existing methods often construct contrastive pairs through random sampling,failing to fully explore the potential of contrastive learning in recommendation systems.To address these issues,this paper introduces a collaborative filtering model based on FeatureNet Contrastive Learning(FCL).The model establishes a node feature similarity matrix by computing the cosine similarity between feature vectors and applying a probabilistic normalization strategy.Using contrastive learning to perform influence analysis on the node feature similarity matrix,the model captures high-order connectivity between nodes,particularly demonstrating significant effectiveness in handling datasets with high sparsity.Extensive experiments conducted on multiple datasets prove the effectiveness of the proposed model.
Study on Graph Data Augmentation Based on Graph Entropy Theory
FU Kun, CUI Jingyuan, DANG Xing, CHENG Xiao, YING Shicong, LI Jianwei
Computer Science. 2025, 52 (5): 149-160.  doi:10.11896/jsjkx.240200016
Abstract PDF(2665KB) ( 74 )   
References | Related Articles | Metrics
Graph data augmentation,as a technique aiming to enhance the performance of graph neural networks,involves transforming and expanding the graph structure and node features to increase the diversity and quantity of training data.The integrity of information structures,the smoothness of feature manifold,the diversity of graph,and local dependencies are difficult to comprehensively considered in graph data augmentation.Additionally,over-smoothing and over-fitting problems exist in the training of graph neural networks,which limit their learning capabilities.To address these issues,a graph data augmentation model(NRGE) based on the entropy theory in thermodynamics is proposed.Firstly,a novel definition of graph entropy is introduced to measure the smoothness of the feature manifold.A new data augmentation strategy,whose main idea is to reduce the loss of graph entropy is proposed to generate more appropriate training data.Secondly, the sampling neighbors of the nodes are augmented to ensure the consistency of data augmentation.To increase the diversity of data augmentation,the first-order neighbors of nodes are randomly replaced with their second-order neighbors.Finally,a neighbor-constrained regularization method is introduced,which improves model performance by enforcing prediction consistency between augmented neighbors.Ablation experiments show that the NRGE model effectively reduces the loss of graph entropy by preserving the information structure of triangles,thereby improving learning effect.Three real datasets are trained by the NRGE model.The obtained low-dimensional representation is applied to node classification.Compared with the baseline methods,the NRGE model achieves a performance improvement of 1.1% on the Cora dataset,0.8% on the Citeseer dataset,and a slight decrease of 0.4% on the Pubmed dataset.The experimental results show that the NRGE model can significantly enhance the performance of graph neural networks and improve the generalization ability.
Cancer Pathogenic Gene Prediction Based on Differential Co-expression Adjacent Network
LI Zhijie, LIAO Xuhong, LI Qinglan, LIU Li
Computer Science. 2025, 52 (5): 161-170.  doi:10.11896/jsjkx.240300110
Abstract PDF(2216KB) ( 74 )   
References | Related Articles | Metrics
Cancer is the first killer of human health.With the rapid development of sequencing technology,a massive amount of cancer gene expression data has been accumulated,and using computational methods to predict pathogenic genes has become a new hotspot in cancer research.However,currently,the prediction of pathogenic genes is mostly based on gene interaction networks,and little consideration is given to the potential connection between local network connections and differential gene expression.In response to the above issues,this paper first utilizes gene expression difference data before and after the disease,calculates the correlation between genes through mutual information,and constructs an adjacency network.Then,a feature vector model is designed for predicting cancer pathogenic genes.Vector features include differential expression information of candidate genes and their neighbors.Cancer-related pathogenic and non pathogenic genes are obtained from public databases such as TCGA,OMIM,and GEO,as well as differential expression data of genes before and after illness,for experiments.Differential expression information of genes and their neighbors in adjacency networks are used for cancer pathogenic gene prediction(DICPG).The experimental results show that the DICPG cancer gene classification model has significant biological significance,and its classification accuracy and AUC performance indicators are superior to similar methods.
Computer Graphics & Multimedia
Restoration of Atmospheric Turbulence-degraded Images Based on Contrastive Learning
MIAO Zhuang, CUI Haoran, ZHANG Qiyang, WANG Jiabao, LI Yang
Computer Science. 2025, 52 (5): 171-178.  doi:10.11896/jsjkx.240200020
Abstract PDF(3218KB) ( 75 )   
References | Related Articles | Metrics
Image degradation caused by atmospheric turbulence seriously affects the performance of downstream computer vision tasks such as object detection and image recognition.Existing deep learning-based image restoration models for atmospheric turbulence degradation have achieved good performance,but have not fully utilized the feature information of the turbulence effect.To improve restoration results,a method for restoring of atmospheric turbulence-degraded images based on contrastive learning is proposed.Aiming at the blurring and distortion caused by atmospheric turbulence degradation,a turbulence mitigation block is designed,which integrates a Transformer-based channel information interaction module and a CNN-based spatial information interaction module to suppress the turbulence interference to the image at both global and local levels.At the same time,contrastive learning is introduced to consider the clear image and the degraded image of atmospheric turbulence as positive and negative samples,to pull the output of the atmospheric turbulence restoration network closer to the positive samples and push it farther away from the negative samples in the feature space,so that feature extraction and image restoration can be performed more efficiently.The proposed method achieves 26.78 dB and 22.42 dB PSNR and 0.790 9 and 0.682 0 SSIM on the synthetic Helen dataset and synthetic Places dataset,respectively,which is the best result compared with the existing five methods,and it is suitable for improving the quality of atmospheric turbulence degradation images.
ECG Signal Denoising Method Based on Stationary Wavelet Transform with Hyperbolic TangentThreshold Function
WANG Haiyong, DING Gufei
Computer Science. 2025, 52 (5): 179-186.  doi:10.11896/jsjkx.240100009
Abstract PDF(2917KB) ( 56 )   
References | Related Articles | Metrics
In the acquisition process of ECG signals,there are various kinds of noise filled in the ECG signals,which will make the ECG signals become difficult to identify,thus affecting the diagnosis of medical personnel.Denoising the ECG signal is an important part of ECG signal research.This paper adopts the technique based on stationary wavelet transform,aiming at the defects of hard threshold and soft threshold in the denoising process of stationary wavalet transform,a hyperbolic tangent function withvariable parameters(SWTaVHT) is proposed for denoising ECG signals.Moreover,in order to prevent the loss of some high frequency information segments in the process of denoising,the R-peak location information assisted correction method is used to better retain useful signal features.In order to evaluate the effectiveness of SWTaVHT,experiments are conducted on the public database MIT-BIH for a comparative study with existing methods.Experimental results show that the signal-to-noise ratio(SNR),root-mean-square error(RMSE) and percentage root-mean-square difference(PRD) after denoising are better compared to the existing methods.The SWTaVHT denoises the ECG data without changing the amplitude of the original signals,which is better than the existing methods.
Open Set Recognition Based on Meta Class Incremental Learning
SUN Jinyong, WANG Xuechun, CAI Guoyong, SHANG Zhiliang
Computer Science. 2025, 52 (5): 187-198.  doi:10.11896/jsjkx.240600162
Abstract PDF(3995KB) ( 60 )   
References | Related Articles | Metrics
Traditional image classification algorithms assume that the world is static and closed,whereas the real world is dyna-mic and open,and new categories and their samples are continually emerging,leading to a decrease in the accuracies of traditional image classification algorithms.To address this problem,researchers proposed open set recognition(OSR) problem for the real world which aims at identifying unknown-class samples while maintaining the classification accuracy for known-class samples.However,existing OSR methods generally neglect the further exploitation of identified unknown-class samples and the unknown class samples are scarce in number,so that the classification model is unable to incrementally learn the knowledge of identified unknown class samples,thereby impairing the accuracy and generalization capability of OSR models.Therefore,this paper proposes an OSR method based on meta-incremental learning to improve the accuracy and generalization of OSR models.This method employs a bi-level optimization mechanism to build an OSR model,and then cluster unknown class samples based on deep learning so that the built OSR model can incrementally learn the knowledge of unknown class samples.Specifically,an OSR model based on bi-level optimization mechanism is constructed and trained with few-shot unknown class samples,in order to enable the OSR model to incrementally learn the knowledge of few-shot unknown class samples.Then,a weight excitation attention method is used to obtain the importance of the OSR model's parameters and update non-critical parameters,thereby reducing the impact of incremental learning on the model's ability to classify known-classes.Additionally,a deep learning-based DBSCAN method is designed to extract features and cluster the identified unknown-class samples.Clustered samples are labeled as the same class and performed incremental learning.While samples that are difficult to cluster are rejected,to avoid the impact of too few unknown-class samples on the model's incremental learning effectiveness.Finally,experimental results on four public datasets show that the proposed method outperforms the mainstream open-set recognition methods on AUROC and F1 scores,and adequately learns the knowledge of identified unknown class samples.
Low Overlap Point Cloud Registration Method Based on Deep Position-aware Transformer
KONG Yu, XIONG Fengguang, ZHANG Zhiqiang, SHEN Chaofan, HU Mingyue
Computer Science. 2025, 52 (5): 199-211.  doi:10.11896/jsjkx.240400172
Abstract PDF(4810KB) ( 52 )   
References | Related Articles | Metrics
In response to the issues such as neglecting the fusion of local geometric embeddings in the feature extraction stage,weak correlation in position-aware information between low overlap point cloud pairs in the feature interaction stage,making it difficult to extract more expressive features and deviation in the transformation solved due to some outlier correspondence in the correspondence generation stage,in this paper,a 3D point cloud low overlap registration method based on deep position-aware Transformer(DeepPAT) is proposed,which follows the local to global matching mechanism.A local feature extraction network based on local geometry information is proposed to extract multi-level features from point cloud.Then,a deep position-aware Transformer(DPAT) module is designed to extract the relevant features and overlap information between low overlap point cloud pairs by learning the geometry and spatial position information of the point cloud itself and across frames,so as to carry out low overlap point cloud matching.Finally,a maximal cliques algorithm adjusted by the feature similarity is designed to reduce the position ambiguity caused by the length consistency and eliminate the outlier correspondences.It can be used as a plug-and-play robust estimator to replace traditional robust estimators such as RANSAC and is fully implemented by Pytorch.Evaluating on the synthetic ModelNet dataset and indoor 3DMatch dataset,the experimental results show that DeepPAT reduces the rotation and translation root mean square error to 3.994 and 0.005 on ModelNet datasets,respectively,and DeepPAT outperformed existing methods by at least 4.3 percentage points and 3.6 percentage points in term of registration recall on 3DMatch and 3DLoMatch benchmarks,respectively.
Improved U-Net Multi-scale Feature Fusion Semantic Segmentation Network for RemoteSensing Images
JIANG Wenwen, XIA Ying
Computer Science. 2025, 52 (5): 212-219.  doi:10.11896/jsjkx.240300137
Abstract PDF(2948KB) ( 79 )   
References | Related Articles | Metrics
High spatial resolution of remote sensing images,the large scale differences of different types of objects,and the imba-lance of categories are the main challenges faced by accurate semantic segmentation tasks.In order to improve the accuracy of semantic segmentation of remote sensing images,this paper proposes an improved U-Net multi-scale feature fusion semantic segmentation network for remote sensing image(Multi-scale Feature Fusion Network,MFFNet).The network is based on the U-Net network and includes a dynamic feature fusion module and a gated attention convolution mix module.Among them,the dynamic feature fusion module replaces the skip connection and improves the feature fusion method of the upsampling layer and the downsampling layer to avoid information loss caused by feature fusion,while improving the fusion effect of shallow features and deep features.Gated attention convolution mix module integrates self-attention,convolution,and gating mechanisms to better capture both local and global information.Comparative experiments and ablation experiments are carried out on Potsdam and Vaihingen.The results show that the mIoU of MFFNet on the two datasets reached 76.95% and 72.93% respectively,effectively improving the semantic segmentation accuracy of remote sensing images.
Hypergraph Convolutional Network with Multi-perspective Topology Refinement forSkeleton-based Action Recognition
HUANG Qian, SU Xinkai, LI Chang, WU Yirui
Computer Science. 2025, 52 (5): 220-226.  doi:10.11896/jsjkx.240600125
Abstract PDF(1725KB) ( 83 )   
References | Related Articles | Metrics
Since the human skeleton is a natural topological structure,graph convolutional networks(GCNs) are widely used for skeleton-based human action recognition.In recent research,skeleton sequences are represented as spatio-temporal graphs and topology graphs are used to model the correlation between human joints.However,current GCN-based methods only focus on pairwise joint relationships and ignore potential high-order relationships beyond pairwise relationships,leading to underutilization of the graph structure of skeleton data.To solve this problem,this paper proposes the concept of hypergraph to represent potential high-order relationships of joints.Since the high-order relationships of joints within each frame in the skeleton sequence may vary,the model dynamically learns the high-order correlations within each frame with the K-NN method and initialize the hypergraph structure using the high-level representation of joints.This hypergraph structure can better learn the high-order relationships between joints as the hyperedges dynamically adjust with the evolution of joint features.In current hypergraph neural networks,hypergraph convolution transforms the hypergraph into a simple graph using the Laplace's transformation and then performs graph convolution.This method does not fully utilize the characteristics of the hypergraph.The proposed hypergraph convolution method better utilizes the relationship between hyperedges and hypernodes in the hypergraph,performing hyperedge graph convolution on each hyperedge to learn the high-order relationships between joints.The second problem with current GCN-based human action recognition methods is that the topology built by GCNs to represent pairwise joint relationships is not dynamic enough,such as using the same topology for all frames in a sample.To fully explore the dynamic correlation between pairwise joints,the frame-wise topology modeling method is proposed to capture correlation between pairwise joints under different frames and channel-level topology modeling method is proposed to capture correlation between different feature types.Finally,a hypergraph convolution network with multi-perspective topology refinement(HyperMTR-GCN) is developedfor skeleton-based action recognition,which has a significant advantage on the NTU RGB+D and NTU RGB+D 120 datasets.Specifically,it improves by 3.7% on the X-sub benchmark of NTU RGB+D and by 5.7% on the X-sub benchmark of NTU RGB+D 120 compared to 2s-AGCN.
Artificial Intelligence
Intelligent Error Correction Model for Chinese Idioms Fused with Fixed-length Seq2Seq Network
HE Chunhui, GE Bin, ZHANG Chong, XU Hao
Computer Science. 2025, 52 (5): 227-234.  doi:10.11896/jsjkx.240400035
Abstract PDF(3321KB) ( 74 )   
References | Related Articles | Metrics
As a special kind of words,four-character idioms are very popular in Chinese.With the development of Chinese error correction task,intelligent error correction for Chinese idioms has become a research hotspot in natural language processing(NLP) domain.For the low accuracy of the existing methods in intelligent error correction task for Chinese idioms,this paper proposes an intelligent error correction model for Chinese idioms fused with fixed-length Seq2Seq network.In the bottom layer,Seq2Seq network architecture and attention mechanism are combined with hybrid dataset construction method to train Seq2Seq model with fixed input and output sequence length,which is used to solve intelligent error correction task for Chinese four-character idioms.Experimental results on a large public Chinese idiom error correction dataset show that the performance of fixed-length Seq2Seq model is better than the existing methods,and it can achieve the goal of intelligent error correction of three diffe-rent Chinese idioms:out-of-order,missing character and wrong character.Its comprehensive error correction accuracy can reach 91.3%,which is 11.73% higher than the optimal baseline model.
Simplification Method for Contradiction Separation Clause in First-order Logic AutomatedTheorem Prover CSE
WU Xin, CHEN Shuwei, JIANG Shipan
Computer Science. 2025, 52 (5): 235-240.  doi:10.11896/jsjkx.241000175
Abstract PDF(1559KB) ( 42 )   
References | Related Articles | Metrics
First-order logic automated theorem proving has the capacity to resolve a multitude of practical problems after formalization,and thus holds considerable practical value.As an advancement in automated theorem proving,contradiction separation deduction extends the classical resolution principle and exhibits enhanced proving capability.In this paper,a simplified algorithm for contradiction separation is proposed and theoretically proven based on the Contradiction Separation Extension(CSE) prover,which follows the rule of contradiction separation.The proposed algorithm enhances efficiency via data structure optimization,utilizing pointers to store complementary information between clauses.This information is then employed to select the clauses that are involved in deductions,thereby achieving the separation clause simplification.This approach produces shorter separation clauses while leveraging the unification complementarity of clauses to strengthen the detection capability of empty clause derivation paths,ultimately improving prover efficiency.Experimental results demonstrate that the enhanced prover CSE_BSCS with this simplification algorithm solves 39 additional test cases compared to the original CSE,with an 18.64% reduction in average proving time. These improvements confirm the superior performance of CSE_BSCS in both proving capability and efficiency over CSE.
Multi-assistant Dynamic Setting Method for Knowledge Distillation
SI Yuehang, CHENG Qing, HUANG Jincai
Computer Science. 2025, 52 (5): 241-247.  doi:10.11896/jsjkx.240700059
Abstract PDF(3029KB) ( 62 )   
References | Related Articles | Metrics
Knowledge distillation is increasingly gaining attention in key areas such as model compression for object recognition.Through in-depth research into the efficiency of knowledge distillation and an analysis of the characteristics of knowledge transfer between the teacher and student models,it is found that the reasonable setting of an assistant model can significantly reduce the performance gap between the teacher and student.However,the unreasonable choice of the scale and number of assistant models can have a negative impact on the student.Therefore,this paper proposes an innovative multi-assistant knowledge distillation training framework,which optimizes the process of knowledge transfer from the teacher to the student by dynamically adjusting the number and scale of assistant models,thereby improving the training accuracy of the student model.In addition,this paper also designs a dynamic stopping strategy for knowledge distillation,sets student models with different training methods as a control group,and achieves personalized design of the stopping rounds for knowledge distillation,further improving the training efficiency of the student model and constructing a more streamlined and efficient multi-assistant knowledge distillation framework.Experiments on public datasets prove the effectiveness of the proposed multi-assistant dynamic setting method for knowledge distillation.
Research on Intelligent Judgment of Criminal Cases Based on Large Language Models
CONG Yingnan, HAN Linrui, MA Jiayu, ZHU Jinqing
Computer Science. 2025, 52 (5): 248-259.  doi:10.11896/jsjkx.241100100
Abstract PDF(2513KB) ( 71 )   
References | Related Articles | Metrics
The intelligentization of criminal case trials has been a hot research topic in the development of digital courts.In the conventional method based on natural language processing,the model directly predicts the final judgment based on the facts of the case.However,when dealing with complex criminal cases,the model may fail to identify the logical dependencies between legal elements and to clearly present the legal reasoning process.The intelligent criminal case trial method based on large language models proposed in this paper follows the approach of “annotating case corpus-pre-training large language model-reinforcing trial logic”.The first step is to annotate the legal elements of the case such as subjects,objects,subjective elements,and objective elements by combining automated annotating with manual correction and create a structured reasoning dataset.The second step is to use ChatGLM3-6b-32k as the foundational large language model for incremental pre-training based on the GLM pre-training framework.The last step is to fine-tune the parameter and increase legal knowledge using the LoRA parameter-efficient fine-tuning strategy and large language model retrieval enhancement technology,thereby reinforcing the trial logic.Experimental results indicate that,compared to Qwen-7B-Chat and Baichuan2-7B-Chat,the ChatGLM3-6b-32k model exhibits superior performance after supervised fine-tuning.The introduction of judicial syllogism significantly enhances the logicality of the judgment texts,ma-king them closer to the reasoning of human judges.In the tasks of charge prediction and sentencing prediction,the model created using this method shows a significant improvement in accuracy compared to the MTL-Fusion,Lawformer,and BERT models.In addition,compared to Legal-BERT and CaseLawBERT,which are trained on European and American legal texts,the ChatGLM3-6b-32k model better suits the trial logic of Chinese criminal cases and demonstrates stronger capabilities in handling long texts.This paper not only explores the application of large language models in intelligent criminal case trials,but also provides valuable references for research on large language models in justice.
Knowledge Graph Completion Method Fusing Entity Descriptions and Topological Structure
HAN Daojun, LI Yunsong, ZHANG Juntao, WANG Zemin
Computer Science. 2025, 52 (5): 260-269.  doi:10.11896/jsjkx.240300012
Abstract PDF(1757KB) ( 55 )   
References | Related Articles | Metrics
Knowledge graph completion aims to predict missing entities and relationships in given triplets to enhance the completeness and quality of the knowledge graph.Existing knowledge graph completion methods typically only consider the structural information of triplets or the individual additional information of entities,such as textual descriptions or topological structure information.This overlooks the fusion of multiple types of additional information to enhance entity feature information,leading to suboptimal performance in completing missing entities.To address this issue,this paper proposes a knowledge graph completion method integrating entity text descriptions and topological structure information,referred to as FuDS-KGC,to enhance the performance of knowledge graph completion tasks.This method first extracts relationship-specific feature representations from entity textual descriptions using Transformer and attention mechanisms to enhance the representation feature information of entity descriptions.Next,it constructs first-order neighbor subgraphs for entities and obtains topological structure features through a graph attention network.Finally,a dynamic gated fusion mechanism is designed to integrate entity textual descriptions and topo-logical structure features to enhance the comprehensive feature representation of entities and overcoming the limitation of existing research focusing on the fusion of singular additional information.Experimental results on FB15k-237 and WN18RR datasets demonstrate the effectiveness of FuDS-KGC.
Finitely-valued Terminal Zeroing Neural Networks with Application to Robotic Motion Planning
WANG Liming, ZHONG Guomin, SUN Mingxuan, HE Xiongxiong
Computer Science. 2025, 52 (5): 270-280.  doi:10.11896/jsjkx.240400173
Abstract PDF(3758KB) ( 63 )   
References | Related Articles | Metrics
To solve the problem of time-varying quadratic programming with equality constraints,this paper proposes a finitely-valued terminal zeroing neural network achieving the finite-time convergence of computing errors while being easy to implement.The convergence of the finitely-valued terminal zeroing neural network is theoretically analyzed,and the specific expression for settling time is provided.The repetitive motion planning problem of redundant robotic manipulators can be described as a time-varying quadratic programming.By employing the finitely-valued terminal zeroing neural network as a solver,the joint position and velocity trajectories,corresponding to the desired end-effector trajectory,can be obtained.Considering the inevitable initial joint shift,a terminal optimization criterion with fixed/adaptive parameters is proposed,aiming for finite-time convergence of end-effector position error and higher precision in repetitive motion planning.To ensure smooth operation of the robotic manipulator,a smooth-corrected finite-value function is proposed for the terminal optimization criterion,and the finite-time convergence of the end-effector position error is established.Numerical simulations and UR5 manipulator simulation and experimental results validate effectiveness of the proposed computing scheme.
Research on Dynamic Incremental Backward Cloud Transformation Algorithm
XU Changlin, KONG Lingzhuo
Computer Science. 2025, 52 (5): 281-290.  doi:10.11896/jsjkx.240100017
Abstract PDF(4454KB) ( 71 )   
References | Related Articles | Metrics
As a tool for studying uncertain information,cloud model is of great significance in uncertain artificial intelligence and data mining.The backward cloud transformation algorithm is one of the important algorithms of cloud model,which can realize the transformation from quantitative data to qualitative concepts.This paper mainly studies the backward cloud transformation algorithm from the perspective of dynamic increment.Firstly,the irrationality of parameter estimation in the existing classical backward cloud transformation algorithm based on the first-order absolute central moment is analyzed theoretically.Secondly,on the basis of theoretical analysis,combined with the characteristics of cloud droplets generated by the forward cloud transformation algorithm,the normal random variable is used to dynamically generate new cloud droplets as new samples,then the randomly generated samples and the original samples are fused as the final samples to estimate the parameters,which effectively solves the estimation problems existing in the existing algorithms.Therefore,two dynamic incremental backward cloud transformation algorithms are proposed.Thirdly,through random simulation experiments,this paper compares the proposed backward cloud transform algorithm with existing algorithms from four aspects:effectiveness,stability,convergence and parameter robustness.The experimental results show that the dynamic incremental backward cloud transformation algorithm proposed in this paper has smaller estimation error,better stability and convergence,and has strong robustness to parameter changes.Finally,the proposed backward cloud transform algorithm is applied to the simulation and evaluation of Shooters' shooting level.The experimental results further show that the proposed algorithms have preferably practicability.
Computer Network
EvoTrace:A Lightweight In-band Network Telemetry Method Based on Nonlinear Embedding and Batch Processing
WANG Panxiang, CUI Yunhe, SHEN Guowei, GUO Chun, CHEN Yi, QIAN Qing
Computer Science. 2025, 52 (5): 291-298.  doi:10.11896/jsjkx.240100164
Abstract PDF(3727KB) ( 51 )   
References | Related Articles | Metrics
In-band network telemetry(INT) enables packets to carry network state information,achieving high monitoring accuracy and precision.However,this advancement comes at the cost of increased data plane overhead.The embedding of telemetry information results in excessive network overhead within the data plane.Meanwhile,existing telemetry methods usually measure large number of packets from large flows,fails in measuring packets from small and medium flows.To address these issues,this paper proposes a lightweight INT method－EvoTrace.EvoTrace introduces a nonlinear packet telemetry method that monitors packets from different flows according to the attributes of network flows.Additionally,EvoTrace also employs a metadata batching method to aggregate the telemetry metadata,for reducing network bandwidth occupancy and the number of telemetry packets.EvoTrace is implemented on OpenvSwitch(OVS) and tested,experimental results demonstrate that,compared with the existing methods,EvoTrace achieves larger network flow monitoring coverage while saves more than 40% of INT bandwidth occupancy.
Study on DApp Resource Naming and Addressing Technology
YE Jueyu, LUN Zhanqun, YUE Qiaoli, LI Hongtao, ZHANG Haikuo, QIANG Jishen
Computer Science. 2025, 52 (5): 299-306.  doi:10.11896/jsjkx.240700205
Abstract PDF(2610KB) ( 48 )   
References | Related Articles | Metrics
In recent years,DApps(Decentralized Application) such as NFT(Non-Fungible Token),DeFi(Decentralized Finance),Metaverse have ignited the enthusiasm for Web3.0 at home and abroad.DApp resources are partially stored in traditional Web servers and partially stored in decentralized systems such as blockchain,and their naming and addressing technology is quite different from that of traditional Web2.0 website resources.However,there is almost no research in related aspects in the academic community.In view of this situation,this paper systematically reviews the existing DApp resource naming and addressing technologies,based on the both of traditional domain names and Web3.0 domain names,and analyzes the advantages and disadvantages of existing technologies.On this basis,aiming at the shortcomings of the existing technology in ease of use,the new DApp resource naming and addressing techniques for DApp are proposed,the functional architecture and implementation framework are designed,and a prototype system is developed for verification.Experimental results show that the proposed new technology has good usability and no significant degradation in performance compared with the existing technology.
Information Security
Survey of Personalized Location Privacy Protection Technologies
CAO Tengfei, YIN Runtian, ZHU Liang, XU Changqiao
Computer Science. 2025, 52 (5): 307-321.  doi:10.11896/jsjkx.240600067
Abstract PDF(2178KB) ( 73 )   
References | Related Articles | Metrics
With the proliferation of mobile networks and smart devices,users' geographical location information is being extensively collected and utilized,posing severe challenges to data privacy.In this context,users not only expect to receive effective security safeguards,but also demand higher quality service experiences.However,protecting users' location privacy often requires limiting or blurring the precision of location information,which conflicts with the high-precision location data needed to provide personalized services.Therefore,how to balance location privacy protection and meeting users' personalized needs has become a critical scientific issue.This issue involves multiple domains such as data security,user experience,and commercial interests,and plays a crucial role in enhancing privacy protection,strengthening user trust,and improving the quality of user service experiences.This paper reviews the recent research progress in personalized location privacy protection.Firstly,it analyzes the causes of privacy breaches and common attack methods.Subsequently,it summarizes the definition and classification of location privacy protection technologies.Then,based on users' personalized needs,it discusses how to provide more suitable location privacy protection measures while ensuring users' privacy preferences.Finally,itsummarizes and looks forward to the future research trends in personalized location privacy protection technologies.
Error Analysis and Parameter Recommendations for Randomness Test Under Large Sample Conditions
SUN Yueyue, FAN Limin
Computer Science. 2025, 52 (5): 322-329.  doi:10.11896/jsjkx.240700006
Abstract PDF(1767KB) ( 38 )   
References | Related Articles | Metrics
In the field of information security,randomness tests play a crucial role in ensuring the security of cryptographic systems.The stability and reliability of these tests directly impact the overall security of cryptographic systems,making error issues during the testing process a focal point for both academia and industry.Particularly when handling large-scale samples,the accumulation of errors can more readily lead to reliability issues in randomness testing.Consequently,studying methods to enhance the accuracy and reliability of randomness testing is of significant importance.The GM/T 0005-2021 standard outlines 9 tests with variable parameters designed for randomness testing of large binary data samples.This study categorizes these tests according to their characteristics and conducts a quantitative error analysis.Specifically,when the bit length of the binary sequence under test is 1×108,the parameters recommended by the GM/T 0005-2021 standard are generally reasonable.For the Maurer universal statistical test,a subsequence length of 6 results in upper bound p-value error of 0.001 492 8,demonstrating higher accuracy compared to the parameters suggested in the GM/T 0005-2021 standard.Similarly,for the linear complexity test,using smaller subsequence lengths results in smaller errors. With the increase in sample length,this study extends the analysis to parameter selection for a sample length of 1×109.It systematically examines the errors associated with different sample lengths and parameter configurations,providing refined parameter recommendations for randomness testing when the sample length reaches 1×109.
Study on Security Risk Relation Extraction Based on Multi-view IB
LI Xiwang, CAO Peisong, WU Yuying, GUO Shuming, SHE Wei
Computer Science. 2025, 52 (5): 330-336.  doi:10.11896/jsjkx.240300162
Abstract PDF(2257KB) ( 43 )   
References | Related Articles | Metrics
Safety risk management is the core assignment to ensure safety,and the traditional methods of identifying safety risks can no longer meet the needs of intelligent development.Research on relation extraction is of significant importance for security risk management,as it serves as one of the methods for identifying security risks.However,most existing relation extraction models ignore the problem of insufficient representation of domain entity and contain more noise in the data.To address the above problems,a multi-view IB-based safety risk relation extraction model(MIBRE) is proposed.Specifically,it achieves enhanced domain entity semantics by fusing semantic information from multi-view.In order to obtain the maximum relevant information between the two views,an objective function is constructed using the information bottleneck approach.The relevant information is maximally preserved and restored while compressing the information between the two views.Experiments on two real domain datasets show that the F1 value recognized by MIBRE reaches 64.28% and 74.34% respectively,which is 4.41% and 2.98% higher than that of LGGCN based on heterogeneous graph model.Compared with TDGAT based on attention mechanism model,F1 value increased by 1.89% and 1.53% respectively.The effectiveness of the proposed model in security risk identification is verified by experiments.
Blockchain-based Internet of Things Traceable and Anonymous Cross-domain AuthenticationScheme
WANG Qiuli, REN Zhiyu, WU Xiangyu, GUAN Qiuguo, WANG Haichao
Computer Science. 2025, 52 (5): 337-344.  doi:10.11896/jsjkx.240100190
Abstract PDF(2103KB) ( 58 )   
References | Related Articles | Metrics
With the wide application of Internet of things technology,there is an increasing demand for cross-domain information sharing,and cross-domain authentication scheme is the foundation for ensuring cross-domain secure collaboration.Realizing cross-domain authentication based on the real identity of the device has the risk of privacy leakage,while the anonymous authentication scheme has the hidden danger of making it difficult to track malicious devices.To address the above problems,a traceable and anonymous cross-domain authentication scheme based on blockchain technology is proposed.Combining one-way hash chain and certificateless cryptography,multiple unrelated pseudonym identities and corresponding public-private key pairs are generated for the device.Dynamic accumulator is used to calculate the changed domain information.Different pseudonyms are used for each cross-domain authentication,and identity authentication is performed based on the domain information and the cross-domain credentials issued by the key generation center,which not only protects the privacy of the device,but also recovers the real identity of the malicious device and holds them accountable.BAN Logic Correctness analysis and formal security proofs show that the proposed scheme has high security.Compared with other schemes,the calculation cost and communication cost in authentication process are lower.
Remote Dynamic Data Integrity Checking Scheme for Multi-cloud and Multi-replica
TAN Shiyi, WANG Huaqun
Computer Science. 2025, 52 (5): 345-356.  doi:10.11896/jsjkx.240300027
Abstract PDF(2271KB) ( 48 )   
References | Related Articles | Metrics
More and more data owners would like to store their data to cloud servers in order to reduce their localstorage burden along with rapid development of cloud servers.However,data owners will lose the direct control over their data after uploading to cloud servers.Data integrity checking is essential to ensure the integrity of remote data stored on cloud servers.It allows data owners to verify the integrity of the outsourced data without downloading all the data.To improve the availability and durability of outsourced data,data owners store multiple copies on multiple cloud servers.It is necessary to protect data owners' identity privacy in public cloud environment because public cloud servers are not completely trustworthy.When data owners want to modi-fy the data stored on the cloud servers,data dynamic operations such as data modification,data deletion,and data insertion are of great significance.Therefore,a remote dynamic data integrity checking scheme in a multi-cloud and multi-replica environment is proposed.The scheme combines the ring signature algorithm with a multi-cloud and multi-replica environment to effectively protect the privacy of data owners' identity,so that data owners do not have to worry about the problems due to identity exposure.At the same time,a new data structure,divide and conquer adjacency table,is introduced to implement dynamic operations of data in multi-cloud environment.The divide and conquer adjacency table searches the specified data through indexes and completes the insertion and deletion of data by modifying the pointers,which enhances updating efficiency compared to other data structures such as Merkle tree.The proposed scheme is secure based on the standard difficulty problem.This scheme makes use of identity-based public key cryptosystem and eliminates complex certificate management.Through performance and security analysis,the scheme satisfies unconditional anonymity,dynamics,and remote data integrity verification.
Multi-factor Dummy Location Selection Algorithm in Location-based Service
LI Yongjun, ZHU Yuefei, WU Wei, BAI Lifang
Computer Science. 2025, 52 (5): 357-365.  doi:10.11896/jsjkx.240200067
Abstract PDF(3219KB) ( 52 )   
References | Related Articles | Metrics
In view of the existing dummy location selection methods in LBS snapshot location privacy protection,the background knowledge attack caused by the time factor of the location itself is ignored,and the sensitive locations are treated equally.Based on this,a multi-factor dummy location selection algorithm(MFDLS) is proposed,which comprehensively considers the factors that affect privacy leakage,including background knowledge such as geographical attributes,semantic attributes,time attributes of the location and query probability as well as the users' sensitive preferences.To ensure that the selected dummy locations can not only effectively resist location homogeneity attack,location semantic attack and query probability distribution attack,but also deal with multiple threats such as location distribution attack,sensitive homogeneity attack and link attack.The algorithm selects the dummy locations that meet the requirements of query probability close to the initiating time,semantic diversification,large anonymous space and relatively consistent time,non-outlier and central point.Compared with the existing dummy location selection algorithm,the security analysis and simulation results show that the proposed algorithm improves the adversary error by at least 16% and reduces the quality loss by at least 30%,which can more effectively resist the background knowledge attack and meet the users' privacy requirements.
BDBFT:A Consensus Protocol Based on Reputation Prediction Model for IoT Scenario
WANG Pu, GAO Zhanyun, WANG Zhenfei, SONG Zheli
Computer Science. 2025, 52 (5): 366-374.  doi:10.11896/jsjkx.240300018
Abstract PDF(2013KB) ( 49 )   
References | Related Articles | Metrics
Blockchain technology has the advantages of strong data security and high trustworthiness in IoT scenarios,but the consensus algorithm in blockchain technology has the disadvantages of high energy consumption,high computational cost,and low scalability,and the deployment of blockchain system in IoT applications faces the problems of low storage capacity,low energy consumption,and low computational capacity of IoT nodes.Based on practical Byzantine fault tolerant algorithm(PBFT),this paper proposes a consensus protocol based on reputation prediction model(BDBFT) in IoT scenario.Firstly,nodes are grouped according to the geographic location classification criteria of the grouping policy to select consensus nodes and reduce the communication delay of intra-group communication.Secondly,a fine-grained reputation prediction model based on Dirichlet distribution is introduced to dynamically update the model according to the reputation information in the lifecycle of each round of view,and nodes with high prediction probability based on the historical and current reputation information are voted as the consensus nodes.The simulation experiment results show that compared with PBFT algorithm and LRBFT algorithm,BDBFT algorithm effectively reduces the probability of Byzantine nodes participating in the consensus,and has obvious performance improvement in four aspects:time delay,throughput,communication overhead and security.
Intrusion Tolerance Scheduling Algorithm for Microservice Workflow Based on Deep Reinforcement Learning
LI Yuanbo, HU Hongchao, YANG Xiaohan, GUO Wei, LIU Wenyan
Computer Science. 2025, 52 (5): 375-383.  doi:10.11896/jsjkx.240500033
Abstract PDF(2422KB) ( 72 )   
References | Related Articles | Metrics
With the rapid development of microservices and container technology,applications executed in the cloud can be completed by multiple microservices with dependencies.However,microservices for container clouds face many security threats due to shared resources.Attackers in the cloud can destroy them directly or indirectly through side channels,container escape,resulting in incorrect output results,which will bring huge losses to users in the cloud.Therefore,an intrusion tolerance scheduling algorithm for microservice workflow(ITSAMW) is proposed to improve the security of the system under the container clouds.Firstly,ITSAMW builds three replicas of each microservice and uses a voting mechanism to guarantee security.ITSAMW studies how to schedule these microservice replicas and proves the location constraints that microservice intrusion tolerance scheduling needs to meet.Secondly,it constructs a microservices scheduling and completion delay model,redefines the security scheduling problem of microservices,and solves the problem with deep reinforcement learning.Finally,in order to verify the effectiveness of ITSAMW,experiments are conducted by using the container clouds simulation platform that Kubernetes builds and are evaluated by using intrusion tolerance,completion delay and load balancing.Experimental results show that compared with the existing methods,under the condition that the completion delay of ITSAMW is increased by 17.6%,the intrusion tolerance is increased by 28.1%,and the load balancing is reduced by 13.7%.
Reversible Facial Privacy Protection Method Based on “Invisible Masks”
ZHENG Xu, HUANG Xiangjie, YANG Yang
Computer Science. 2025, 52 (5): 384-391.  doi:10.11896/jsjkx.241100066
Abstract PDF(3052KB) ( 53 )   
References | Related Articles | Metrics
With the rapid progress of artificial intelligence and computer vision technology,facial information has been widely used in smart security,financial payment,and social media,etc.Once the collected facial information is leaked or illegally sold by unscrupulous individuals,it will cause adverse consequences.Therefore,how to prevent the original facial database from being illegally accessed and trained by malicious parties,and how to prevent illegal recognition,is an urgent issue that needs to be solved.Therefore,a reversible facial privacy protection method based on “invisible mask” is proposed.If the adversarial facial image is illegally accessed,it will cause the unauthorized facial recognition system to incorrectly recognize,and for authorized users,the original facial information can be recovered by removing the “invisible mask”,ensuring that the authorized facial recognition system can correctly recognize,thus achieving the purpose of protecting the facial database.Experimental results show that the method generates adversarial facial images with higher visual quality,the average PSNR between the adversarial facial image and the original facial image without attack layer can reach 55 dB,and the false recognition rate of the unauthorized system can reach 99.6%.At the same time,the method realizes reversible recovery of facial images,the average PSNR of the recovered facial image is 61 dB,and the correct recognition rate of the authorized system can reach 99.8%.Therefore,the proposed method can effectively protect the facial database.
渝ICP备16013121号-4
Add:18# Honghu West Rd., Yubei ,Chongqing,China    Post Code: 401121
Tel: 023-63500828/023-67039612/023-67039623
E-mail: jsjkx12@163.com
Powered by Beijing Magtech Co., Ltd.