Computer Science

Survey of Tabular Data Generation Techniques

WANG Yongxin, XU Xin, ZHU Hongbin

Computer Science. 2025, 52 (10): 3-12. doi:10.11896/jsjkx.250800044

Abstract

PDF(2467KB) ( 9 )

References | Related Articles | Metrics

Tabular data holds significant value due to its widespread application in critical domains such as finance and healthcare.However,the effective utilization of tabular data is often constrained by data scarcity,class imbalance,and stringent privacy regulations.To address these challenges,synthesizing samples that are statistically highly similar to real data through generative models has emerged as a novel solution,aiming to enhance data availability and protect user privacy.The technological development path in this field has progressively evolved from traditional deep learning models to cutting-edge paradigms.Early explorations are represented by Variational Autoencoders and Generative Adversarial Networks,but these methods often face bottlenecks such as training instability and mode collapse,affecting the quality of generated data.To overcome these difficulties,diffusion models have emerged,demonstrating significant advantages in generating high-fidelity and diverse samples through a progressive denoising process.Nevertheless,the core of these models remains the imitation of statistical distributions,lacking an understanding of real-world common sense.Consequently,the latest research has shifted towards methods based on Large Language Models(LLMs),leveraging their rich world knowledge to generate synthetic tabular data that is not only statistically authentic but also logically and semantically more reasonable.A systematic review of this field aims to provide researchers and practitioners with a comprehensive understanding of the technology and offer decision-making references for selecting the most appropriate technical path in different application scenarios.

Research on Portfolio Construction Based on Topological Structure Features

LI Ruiyang, LI Shuyi, YANG Yuexi, PENG Chuhan, XING Jingyu, QIAO Gaoxiu

Computer Science. 2025, 52 (10): 13-21. doi:10.11896/jsjkx.250100136

Abstract

PDF(3299KB) ( 8 )

References | Related Articles | Metrics

In recent years,the application of topological data analysis(TDA) in the financial field has gradually demonstrated its value.TDA,through methods such as persistent homology,constructing complexes that effectively quantify the shape of data,facilitating the extraction of data information.This provides unique advantages for time series analysis,particularly in the clustering of financial time series and the construction of portfolios.Based on this,by deeply mining the time series data of China's stock market using TDA methods,combined with clustering algorithms,and applying these insights to portfolio construction,the effectiveness of such approaches is analyzed.The results,validated through the sliding window method,indicate that portfolios constructed based on TDA(denoising) clustering perform well in terms of return-risk ratio and stability,outperforming the overall market.Therefore,the TDA method can more effectively mine information from stock data,providing a scientific basis for investors to optimize returns.

Group Cross Adversarial Application in Stock Price Prediction

LI Ao, BAI Xueru, JIANG Jiali, QIAO Ye

Computer Science. 2025, 52 (10): 22-32. doi:10.11896/jsjkx.250300104

Abstract

PDF(3390KB) ( 5 )

References | Related Articles | Metrics

Stock price prediction has always been a key topic of focus in both financial research and quantitative investment.To address the issues of mode collapse and weak generalization ability in traditional GAN models,and to improve stock price prediction accuracy,this study proposes a Group Cross-Adversarial model (GCA).The model includes multiple generators and multiple discriminators,and introduces a collaboration mechanism between the generators and discriminators to enhance the generalization ability of the generators.Additionally,knowledge distillation is used to further improve the prediction performance of the generators.The experiment selects daily data of 8 stocks (Industrial and Commercial Bank of China,Huaneng International,China Merchants Bank,and Qingdao Haier from the A-share market,and Alibaba,Amazon,JD,and Bank of America from the US stock market) from January 1,2015 to January 1,2025 as the research sample,and constructs a dataset of 24 feature variables,including market data and technical indicators.The results show that the proposed GCA model significantly outperforms the standalone GRU,LSTM,and Transformer models in terms of the three evaluation metrics－MAE,MAPE,and MSE.Additionally,it surpasses the GRU-GAN,LSTM-GAN,and Transformer-GAN models,which integrate GAN architectures,as well as the WGAN-GP and ResNLS models.Even without optimizing the original models with GAN,the introduction of the GCA framework still improves the prediction accuracy.Further discussion indicates that increasing the number of generator-discriminator pairs can further optimize prediction performance.

Integration of Machine Learning Prediction and Water Wave Optimization for Online Customer Service Representatives Scheduling in Bank Contact Centers

LU Xueqin, XIE Xicheng, TANG Yan, CHEN Shikun, LIU Yangguang

Computer Science. 2025, 52 (10): 33-49. doi:10.11896/jsjkx.250500086

Abstract

PDF(5164KB) ( 1 )

References | Related Articles | Metrics

Online customer service representative scheduling is a crucial component of operational management in a bank's contact center.Optimized staffing and shift scheduling of customer service representatives ensure prompt customer responses,which significantly improves service efficiency and customer satisfaction.However,factors such as the uncertainty of customer request arrivals and variations in customer service skill levels make the online service representatives scheduling problem complex.Considering the practical challenges such as customer service representative skill levels,diversity of customer types,and matching requirements,this study proposes a mixed-integer linear programming model with the optimization objectives of minimizing custo-mer waiting time and operational costs,and also presents a hybrid machine learning and water wave optimization(WWO) method to effectively solve the online customer service scheduling problem.In this method,a forecasting model based on long short-term memory neural networks is employed to predict customer arrival volumes,and this model can capture both time series dependencies and the influence of external factors.For the mixed-integer programming model of representative scheduling,WWO combining reinforcement learning Q-learning is used for efficient solution.This method leverages Q-learning to adaptively optimize neighborhood selection,enhancing the efficiency and quality of solutions.Based on real data from the contact center of Zhejiang Tailong Bank's Ningbo branch,the experimental results show that the proposed method significantly outperforms comparative methods in terms of operational cost control.Further,sensitivity analysis reveals that when forecast accuracy drops below 90%,customer arrival uncertainty markedly increases scheduling costs and customer waiting times.Conversely,when accuracy reaches or exceeds 90%,performance improvements stabilize.These findings highlight the critical role of high-precision forecasting in effective scheduling and provide theoretical insights and practical guidance for balancing forecasting model complexity with scheduling efficiency in real-world applications.

Interpretable Credit Risk Assessment Model:Rule Extraction Approach Based on AttentionMechanism

WANG Baocai, WU Guowei

Computer Science. 2025, 52 (10): 50-59. doi:10.11896/jsjkx.250300059

Abstract

PDF(2439KB) ( 2 )

References | Related Articles | Metrics

Due to the limitations of traditional statistical models for credit risk assessment,machine learning techniques have significantly enhanced model accuracy and predictive capabilities.However,the complexity and opacity pose significant challenges in terms of interpretation.To address this issue,this paper introduces an interpretable machine learning model for credit risk assessment that integrates the attention mechanism with tree ensemble rule extraction approach.This model automatically identifies complex nonlinear relationships within the data,extracts a large number of interpretable rules from the trained tree ensemble model,encodes these rules into new feature variables,and inputs them into an attention neural network to obtain attention weights for each rule.Subsequently,based on the attention weights,objective function,and constraints,the model balances the predictive performance,stability,and interpretability of the rule subset.The optimal rule subset can be derived in O(n) time.Experimental results,based on three public datasets,demonstrate that the proposed approach not only maintains high predictive accuracy but also substantially enhances the model's interpretability.

Anti-money Laundering Detection Method for Asset Management Based on Temporal Graph Neural Networks

XU Xin, ZHU Hongbin, CHEN Jie, LI Qingwen, ZHANG Xiaorong, LYU Zhihui

Computer Science. 2025, 52 (10): 60-69. doi:10.11896/jsjkx.250800009

Abstract

PDF(3104KB) ( 4 )

References | Related Articles | Metrics

The asset management industry,characterized by high-frequency and flexible financial operations,has become a primary target for money laundering activities.However,the sparsity of transaction structures,the complexity of implicit fund transfer paths between accounts,and the non-uniform characteristics of transaction behaviors pose significant challenges for traditional graph modeling methods based on explicit relationships.To address these issues,this paper proposes an Anti-money Laundering Detection Framework for Asset Management based on Temporal Graph Neural Networks(AM-GAML).By integrating temporal models with graph neural networks,the proposed framework constructs a joint temporal-structural embedding representation and designs a graph generation mechanism based on implicit interaction relationships.This enables the framework to effectively capture weakly correlated features in transaction records and uncover complex transaction behavior patterns between users.Experimental results on a real-world transaction dataset demonstrate that AM-GAML significantly outperforms several advanced approaches in terms of accuracy,recall,F1-score,and AUPRC.The framework excels particularly in minority class detection and generalization ability.The proposed method provides an efficient and reliable solution for anti-money laundering detection in the asset management industry and offers valuable support for risk prevention and control research in complex financial scenarios.

Urban Flow Prediction Method Based on Structural Causal Model

LIU Yuting, GU Jingjing, ZHOU Qiang

Computer Science. 2025, 52 (10): 70-78. doi:10.11896/jsjkx.241000088

Abstract

PDF(3985KB) ( 0 )

References | Related Articles | Metrics

Urban flow prediction plays a critical role in smart city research,providing essential data for urban planning and resource optimization.In recent years,Graph Neural Network (GNN)-based models have significantly enhanced the accuracy of urban flow prediction.However,most existing studies assume that training and testing data come from the same distribution,ignoring the complexity of dynamic changes in urban flow distribution in the real world,leading to poor model performance.To address this challenge,this paper proposes an urban flow prediction method based on the structural causal model to effectively deal with the challenge of model generalization caused by distribution shift.This method first utilizes a structural causal model to uncover the impact of environmental factors as confounders on flow prediction.It then designs a shared distribution estimator to learn the prior distribution of environmental information.Furthermore,a backdoor adjustment approach is introduced,combined with variational inference,to effectively eliminate the confounding effects caused by environmental factors.The proposed method can fairly consider different environmental factors,improving the accuracy and robustness of prediction.Experimental results on two real-world datasets show that the proposed model has high prediction accuracy and robustness when dealing with distribution shift.Compared with the six state-of-the-art baselines,the prediction performance is improved by 2.26%~9.18%.

Adaptive Spectral Clustering Algorithm Based on Relative Proximity

YUAN Zefei, ZHANG Zhengjun, JIANG Guolin

Computer Science. 2025, 52 (10): 79-89. doi:10.11896/jsjkx.240800102

Abstract

PDF(4244KB) ( 0 )

References | Related Articles | Metrics

The traditional spectral clustering algorithm with Gaussian kernel function as the similarity measure has the problem that the scale parameter needs to be artificially set,and the similarity is not related to the sample distribution structure.In order to solve this problem,the shared neighbors based on the natural k-nearest neighbors are defined,and a multi-scale parameter reflecting the regional density is constructed based on the nearest neighbors information of the data points,and the similarity mea-sure is redefined with the new scale parameter.This paper proposes an adaptive spectral clustering algorithm based on relative proximity(RPASC).The improved scale parameter combines the characteristics of interval scale,sequence scale and proportional scale,embodying the relative position relationship between data points and reflecting the distribution characteristics and spatial structure of different density clusters,which improves the adaptability of the algorithm to different distribution datasets.The new similarity measure adaptively reduces the similarity of data points on cluster boundaries of different densities by flexibly adjusting the size of local scale parameter,making cluster boundaries more explicit,which is conducive to discovering the true cluster morphologies.Experiments on synthetic datasets and UCI real datasets verify the effectiveness of the RPASC algorithm on multiple clustering performance indicators.

Spatial-Temporal Propagation Graph Neural Network for Traffic Prediction

LEI Ershuai, YU Suping, FAN Hong, XU Wujun

Computer Science. 2025, 52 (10): 90-97. doi:10.11896/jsjkx.241000045

Abstract

PDF(3178KB) ( 3 )

References | Related Articles | Metrics

In the field of traffic prediction,there are complex and long-range spatial-temporal relationships between data.The existing graph structures fail to fully explore the implicit spatial-temporal relationships between data.To address the above problems,this study makes a series of improvements on the multivariate time series forecasting with graph neural networks(MTGNN),and proposes a spatial-temporal propagation graph neural network(STPGNN) for traffic prediction.Firstly,it captures information at different time scales through the multi-scale convolution module,and uses the feature fusion module to fuse to capture complex temporal information.After that,on the basis of the unidirectional adaptive graph structure of MTGNN,a bidirectional graph learning layer is designed and added to deeply explore and utilize the implicit bidirectional spatial relationship between data.Next,for the information transfer between network layers,a new information transfer method is proposed,which transmits the multi-scale time information in each layer to the next layer in turn,so as to better explore the complex and long-range spatial-temporal relationship.Finally,according to the temporal and spatial information of each level of the network,the prediction results are obtained by output convolution.Experiments are carried out on datasets such as METR-LA,PEMS-BAY and NE-BJ.The results show that STPGNN can effectively improve prediction accuracy and is better than some existing methods in three commonly indicators.This is especially true when it comes to longer-term forecasts.

ACCF:Time Prediction Mechanism-driven Top-k Flow Measurement

HU Yongqing, YANG Han, LIU Ziyuan, QING Guangjun, DAI Qinglong

Computer Science. 2025, 52 (10): 98-105. doi:10.11896/jsjkx.241000033

Abstract

PDF(2196KB) ( 4 )

References | Related Articles | Metrics

In addressing the problem that current top-k flow measurement filtering algorithms depend on fixed counter thresholds,a measurement structure named ACCF based on the activity prediction mechanism has been put forward.ACCF incorporates the activity prediction mechanism,utilizing time series analysis and the EWMA mechanism,to dynamically compute the activity of network flows and accomplish real-time identification and early filtering of potential top-k flows.With respect to the accuracy loss that may be induced by hash conflicts,ACCF introduces a SRST for storing the network flow information on the evicted paths.When the eviction operation reaches the predefined MaxNumKicks value,SRST will give priority to evicting the network flow item with the lowest activity within the local scope to avoid the loss of crucial traffic information.Experimental results indicate that,under suitable parameter combinations,ACCF and SRST can filter out approximately 65% of the major flows in advance and reduce insertion operations by approximately 41%,significantly improving the accuracy in top-k traffic measurement,especially when compared with traditional algorithms such as Space Saving (SS),CM Sketch,LUSketch,and Cuckoo Counter,thereby de-monstrating distinct advantages.

Spatial-Temporal Joint Mapping for Skeleton-based Action Recognition

ZHAO Chen, PENG Jian, HUANG Junhao

Computer Science. 2025, 52 (10): 106-114. doi:10.11896/jsjkx.240800108

Abstract

PDF(3494KB) ( 2 )

References | Related Articles | Metrics

In recent years,skeleton-based motion recognition tasks have received extensive attention from researchers and have made great progress in research.As powerful and effective model paradigms,graph convolutional networks and convolutional neural networks are also favored by researchers in the field of skeleton action recognition.However,1)most GCN-based methods use the paradigm of modeling spatial features and temporal features alternately,which hinders the direct communication of spatial-temporal information;2)For CNN-based methods,they effectively model spatial-temporal information.However,compared with GCN-based methods,they do not make good use of spatial information.In order to solve the above problems,this paper proposes a novel method called Spatial-Temporal Joint Mapping(STJM).The proposed method not only combines the topological information of the graph in GCN-based methods,but also uses CNN-based methods to aggregate spatial-temporal information simulta-neously.Compared with the traditional GCN method, the STJM maps the nodes in high dimension and has stronger ideographic ability.After high-dimensional mapping of nodes,only a simple τ×K convolution kernel is needed to aggregate both temporal and spatial features.As a novel spatial-temporal information aggregation module,many GCN-based topology enhancement strategies can be applied to STJM block.Compared with the previous spatial-temporal simultaneous aggregation model,the proposed me-thod has better performance.Experiments show that combining the proposed STJM Block as a plug-and-play module with GCN exceeds the previous state-of-the-art models on two large-scale datasets:NTU RGB+D 60 and NTU RGB+D 120.

Direct PET to CT Attenuation Correction Algorithm Based on Imaging Slice Continuity

ZHENG Hanyuan, GE Rongjun, HE Shengji, LI Nan

Computer Science. 2025, 52 (10): 115-122. doi:10.11896/jsjkx.240700135

Abstract

PDF(3915KB) ( 0 )

References | Related Articles | Metrics

PET attenuation correction technology is of significant clinical importance,effectively reducing cancer misdiagnosis rates and enabling more precise treatment planning.However,traditional PET attenuation correction methods face challenges such as long scan times and errors introduced during post-processing,limiting their applicability.Recently,attenuation correction methods based on directly generating CT from PET have gained popularity in clinical settings due to shorter scan times and the advantage of avoiding post-processing errors.However,the substantial semantic differences and misalignment between PET and CT pose significant challenges for direct PET-to-CT attenuation methods in modal generation.To address this challenge,this paper proposes a PET attenuation correction method based on Cycle-S2SCT-Net.Cycle-S2SCT-Net utilizes a cyclic generative adversarial structure to learn the mapping between PET and CT distributions,facilitating semantic translation between the two modalities.Within a single generative adversarial network,Cycle-S2SCT-Net integrates an imaging slice continuity module to enhance the network's semantic alignment capability,thereby improving the continuity and accuracy of generated images.Addi-tionally,this paper introduces a network feature layer loss function(Layer Loss) to enhance the feature extraction capability of the generation network.The experimental results demonstrate that CT generated by Cycle-S2SCT-Net and its attenuation-corrected PET exhibit excellent performance in both quantitative evaluation metrics,such as peak signal-to-noise ratio(PSNR),structural similarity index(SSIM),root mean square error(RMSE),and visual quality.

SAM-Retina:Arteriovenous Segmentation in Dual-modal Retinal Image Based on SAM

XU Hengyu, CHEN Kun, XU Lin, SUN Mingzhai, LU Zhou

Computer Science. 2025, 52 (10): 123-133. doi:10.11896/jsjkx.240800013

Abstract

PDF(4437KB) ( 2 )

References | Related Articles | Metrics

The shapes of arteries and veins are highly similar in RGB retinal imaging,and their inherent structures are both subtle and complex,making it difficult for most retinal image processing models to achieve ideal results.To improve the accuracy of arteriovenous segmentation and reduce training costs,a retinal segmentation model based on segment anything model is proposed—SAM-Retina.SAM-Retina adopts a feature fusion,adaptive image encoder,and mask decoder architecture,using structure-and-function dual-modal retinal images simultaneously containing RGB as well as 570 nm and 610 nm single wavelength images instead of a single mode(RGB) image as input.The features of these three images are fused through a feature fusion.While retaining the pre-trained parameters of the image encoder on a large-scale natural image dataset,the model's feature extraction capability on retinal medical images is enhanced by inserting an adapter module and updating it within the vision transformer(ViT) block.The static prompt embedding instead of prompt encoder is adopted to remove the input and encoding process of prompts in the original SAM segmentation process.During the experimental phase,the model is trained and evaluated on the DualModal2019 and HRF datasets,and compared with U-Net,CRU-Net,and TW-GAN.The experimental results show that SAM-Retina is more advanced than other models in various evaluation indicators and the employment of dual-modal image also improves segmentation perfor-mance without increasing the model size.

Sparsity Cooperated Correntropy Based Robust Principal Component Analysis

CHEN Ping, LIU Kehan, LIANG Zhengyou, HU Qixing, ZHANG Yuanpeng

Computer Science. 2025, 52 (10): 134-143. doi:10.11896/jsjkx.240800076

Abstract

PDF(12824KB) ( 3 )

References | Related Articles | Metrics

PCA is widely used in many applications but is sensitive to non-Gaussian noise.Many Robust PCA models have been proposed to handle this issue.However,these methods only can handle one type of noise,such as the impulse noise in the feature domain or the outliers in the sample domain.This paper proposes a novel RPCA model based on sparsity cooperated correntropy called SCPCA,which is robust against impulse noise and outlier simultaneously.Furthermore,an iterative algorithm is proposed to solve the proposed model based on the Fenchel conjugate and the accelerated BCU strategy.Extensive experiments on clustering,background reconstruction and face modelling have been conducted to evaluate the robustness of the proposed method.The results show that the proposed method outperforms the compared state of-the-art methods in most situations.

Target Tracking Method Based on Cross Scale Fusion of Features and Trajectory Prompts

WEN Jing, ZHANG Songsong, LI Xufeng

Computer Science. 2025, 52 (10): 144-150. doi:10.11896/jsjkx.240800159

Abstract

PDF(3950KB) ( 0 )

References | Related Articles | Metrics

When Transformer is used alone for feature extraction in object tracking,the absence of inductive bias makes it difficult to adapt to change in target scale and appearance.To address this,this paper introduces target tracking method based on cross scale fusion of features and trajectory prompts(Cross Scale Fusion of features and Trajectory Prompts Tracker CSFTP-Tracker).In constructing the input for the object tracking network,both the template image and the search image are simultaneously fed into an encoder that fuses CNN and ViT.A key design element is the multi-level spatial-aware pyramid module (Multi-Level Spatial Awareness Pyramid,MSAP).Firstly,the multi-scale CNN features are enhanced with self-attention to strengthen target location information.These multi-scale features are then fused with the F-embeddings features from the ViT and input into the ViT encoder.This fusion strategy not only enhances information interaction between patches within the ViT but also enables the network to leverage both the local features of CNN and the global dependency capabilities of the Transformer.Furthermore,the fused features extracted by the ViT,along with the trajectory prompt features,are fed into the decoder,where autoregressive learning is employed to predict the target's position.Experimental results on the GOT-10k dataset show that,compared to the baseline models,the proposed network improves the average overlap(AO) by 1.3% and increases the success rate score at a 0.5 threshold(SR0.5) by 1.4%.

Ship Detection Method for SAR Images Based on Small Target Feature Enhanced RT-DETR

ZHANG Hongsen, WU Wei, XU Jian, WU Fei, JI Yimu

Computer Science. 2025, 52 (10): 151-158. doi:10.11896/jsjkx.250100097

Abstract

PDF(3003KB) ( 1 )

References | Related Articles | Metrics

In ship detection tasks,SAR images are widely used in maritime resource management,search and rescue,and other scenarios due to their excellent imaging conditions.However,traditional target detection algorithms perform poorly due to issues such as the small size of ships and sea surface clutter.Recently,many algorithms have introduced the attention mechanism of Transformer to achieve better semantic interpretation or adopted more complex network structures to improve feature extraction capabilities.This has improved detection accuracy to some extent but has sacrificed detection speed.This paper proposes a ship detection method for SAR images based on small target feature enhanced RT-DETR.The method consists of three parts:1)Large model prompt generation network:Leveraging the zero-shot learning capability of multimodal large models,prompts are generated to extract more discriminative information from the image modality;2)AIFI-EAA module:Using RT-DETR as the baseline,the scale-invariant feature interaction module is improved by introducing an efficient additive attention mechanism to reduce the computational complexity of the algorithm;3)Lightweight small target feature enhancement fusion network:A small target detection layer is added to the multi-scale feature fusion network,and the CSP-OmniKernel module is designed for multi-scale feature fusion to enhance small target detection performance.Experiments on three public datasets(SSDD,HRSID,and SAR-Ship-Dataset) demonstrate that the proposed method has advantages in terms of accuracy.

Immediate Generation Algorithm of High-fidelity Head Avatars Based on NeRF

SHENG Xiaomeng, ZHAO Junli, WANG Guodong, WANG Yang

Computer Science. 2025, 52 (10): 159-167. doi:10.11896/jsjkx.241000066

Abstract

PDF(2791KB) ( 1 )

References | Related Articles | Metrics

To resolve the escalating need for quickly generating high-precision digital humans in fields such as digital entertainment,virtual reality,and the metaverse,this paper proposes a novel method for rapidly generating a high-precision face model based on monocular RGB videos.Meanwhile,a new framework dedicated to precise modeling of the facial and neck regions is constructed.In particular,the proposed framework integrates neural primitives into a parameterized model of the head and neck,utilizing Head-And-neCK(Hereinafter referred to as HACK) as a superior alternative to the widely adopted Face Latent Animated Mesh Estimator(Hereinafter referred to as FLAME).This substitution markedly enhances the precision and efficiency of 3D facial reconstruction.Additionally,the proposed method has designed a real-time adaptive neural radiance field that significantly accelerates the training and reconstruction processes.By introducing a multi-resolution hash grid and employing the nearest triangle search for deformation gradient calculation within the deformation space,the proposed method achieves rapid reconstruction of high-fidelity head and neck models within minutes.Extensive quantitative and qualitative evaluations demonstrate that the proposed model exhibits notable improvements in both rendering quality and training time compared to existing state-of-the-art methods.

Low Light Image Adaptive Enhancement Algorithm Based on Retinex Theory

ZHENG Dichen, HE Jikai, LIU Yi, GAO Fan, ZHANG Dengyin

Computer Science. 2025, 52 (10): 168-175. doi:10.11896/jsjkx.240800057

Abstract

PDF(4125KB) ( 2 )

References | Related Articles | Metrics

Images in real-world environments are often shot under sub-optimal lighting conditions,resulting in insufficient brightness and poor visual experience.Existing low-light image enhancement methods are often complex in structure and focus on improving the visibility of dark areas,but may over-enhance the bright areas of the image and amplify hidden noise.Most methods based on Retinex theory have problems such as excessive noise,loss of details and color distortion,which affect the visual quality of the enhancement results.In order to solve this problem,this paper proposes a low-light image adaptive enhancement algorithm based on Retinex theory,which aims to effectively improve the brightness of the image while restoring the image truly and naturally.Firstly,the low-light image is passed through the projection module to remove noise that is not suitable for Retinex decomposition.Secondly,the decomposition network decomposes the image into an illumination component and a reflection component.Finally,the illumination component is adjusted through an adaptive iterative curve and multiplied with the reflection component to obtain an enhanced image.Experimental results show that compared with other mainstream algorithms,the proposed algorithm has obvious advantages in objective evaluation indicators,especially PSNR and SSIM:tests on the LOL dataset show that PSNR and SSIM reach 19.98 dB and 0.76,respectively,which are 4.9% and 4.1% higher than the suboptimal algorithm;tests on the LSRW dataset show that PSNR and SSIM reach 18.94 dB and 0.58,respectively,which are 1.5% and 7.4% higher than the suboptimal algorithm.On both of the referenced dataset and the non-reference dataset,the brightness of the enhanced image obtained by the proposed algorithm is significantly improved,the colors are more realistic and natural,and the subjective visual effect is better.

Review of Research on Agent Training Methods Toward Human-Agent Collaboration

HUANG Weiye, CHEN Xiliang, LAI Jun

Computer Science. 2025, 52 (10): 176-189. doi:10.11896/jsjkx.241000047

Abstract

PDF(2488KB) ( 4 )

References | Related Articles | Metrics

Human-agent collaboration has received widespread attention in recent years,and multi-agent reinforcement learning has demonstrated significant advantages and application potential in the field of human-agent collaboration.This paper first introduces the basic concepts and important models of multi-agent reinforcement learning,and analyzes the advantages of multi-agent reinforcement learning in human-agent collaborative tasks,and introduces human-agent collaboration in three types.Secondly,it explores three training paradigms of multi-agent reinforcement learning,including centralized training and centralized execution,decentralized training and decentralized execution,and centralized training and decentralized execution,as well as the applicable scenarios for each training paradigm.Then,in response to the problems faced by agent training methods for human-agent collaboration,such as poor generalization ability,lack of diversity in training partners and inability to better adapt to human partners,it summarizes the research progress on agent training methods for human-agent collaboration from the perspective of whether human data is used or not.Finally,it discusses the application scenarios and future development trends of human-agent collaboration,proposes possible solutions and research directions.

Review of Quantum-inspired Metaheuristic Algorithms and Its Applications

RUAN Ning, LI Chun, MA Haoyue, JIA Yi, LI Tao

Computer Science. 2025, 52 (10): 190-200. doi:10.11896/jsjkx.250500127

Abstract

PDF(1644KB) ( 1 )

References | Related Articles | Metrics

The quantum meta heuristic algorithm is developed by applying quantum computing to the meta-heuristic algorithm.This kind of algorithm is good at solving combinatorial and numerical optimization problems,and has the characteristics of acce-lerated convergence,enhanced exploration and development capabilities,and can obtain higher performance results than traditional meta-heuristic algorithms.This paper mainly summarizes and reviews the theoretical methods and applications of quantum meta-heuristic algorithms.Firstly,this paper expounds the basic concepts and principles of quantum computing,and analyzes the challenging problems that need to be solved urgently in the field of quantum computing.Then,this paper expounds the basic principles of six classical quantum meta-heuristic algorithms,analyzes the latest research progress,gives their advantages and disadvantages in solving domain-specific problems,and demonstrates the application of quantum meta-heuristic algorithms in different disciplines and engineering scenarios.Finally,this paper analyzes and explores the existing problems in the theories and methods of quantum meta-heuristic algorithms,and summarizes the future development direction of the theory and application of quantum meta-heuristic algorithms.

Subject Knowledge Evaluation Method for Language Models Based on Multiple ChoiceQuestions

XIONG Zhuozhi, GU Zhouhong, FENG Hongwei, XIAO Yanghua

Computer Science. 2025, 52 (10): 201-207. doi:10.11896/jsjkx.240800148

Abstract

PDF(3446KB) ( 3 )

References | Related Articles | Metrics

Subject knowledge evaluation methods for pre-trained language models based on Multiple Choice Questions(MCQ) offer rapid,quantitative evaluation of model knowledge.However,their reliability is compromised by irrelevant factors such as option order and length,raising robustness concerns.To address this challenge,an analytical framework for evaluating subject knowledge of pre-trained language models using MCQ is proposed.This framework formalizes MCQ evaluation into two mo-dules:prompting and parsing,systematically investigating the impact of various MCQ evaluation methods on evaluation outcomes.The robustness of different prompting and parsing techniques is analyzed through experiments on Chinese and English subject knowledge evaluation datasets.Based on these findings,a rewriting-enhanced parsing method is introduced that employs pre-trained language models to rewrite model responses,effectively overcoming the limitations of traditional rule-based parsing when handling non-standard replies.By integrating rewriting and rule-based parsing,this approach enhances both answer extraction accuracy and evaluation process robustness,offering a novel and effective strategy for language model evaluation.

Multimodal Information Extraction Fusion Method Based on Dempster-Shafer Theory

WANG Jian, WANG Jingling, ZHANG Ge, WANG Zhangquan, GUO Shiyuan, YU Guiming

Computer Science. 2025, 52 (10): 208-216. doi:10.11896/jsjkx.240200081

Abstract

PDF(2075KB) ( 3 )

References | Related Articles | Metrics

In the past MIE tasks,researchers usually use data layer fusion to train neural network models for MIE.However,due to the heterogeneity among different modalities,this fusion approach can lead to issues such as feature redundancy,incompatibility,and lack of interpretability,which in turn affect the effectiveness of model training.In view of this,this paper proposes a decision-level fusion method based on the DS theory to solve the problems of feature redundancy,incompatibility,and lack of interpretability caused by data layer fusion.The evidence is generated by processing different modal features through neural networks and Dirichlet functions,and after evidence correction and weight assignment,the Shafer fusion rule is utilized to arrive at the final decision.This method effectively improves the accuracy of feature processing and the interpretability of the model.Using accuracy,recall,and F1 score as evaluation metrics,experiments on public and private datasets show an overall performance improvement of 0.22 to 1.87 percentage points compared to existing methods.

SPEAKSMART:Evaluating Empathetic Persuasive Responses by Large Language Models

CHEN Yuyan, JIA Jiyuan, CHANG Jingwen, ZUO Kaiwen, XIAO Yanghua

Computer Science. 2025, 52 (10): 217-230. doi:10.11896/jsjkx.241200055

Abstract

PDF(5717KB) ( 1 )

References | Related Articles | Metrics

In recent years,LLMs have shown amazing capabilities in emotional dialogues and strong goal-achievement abilities.However,existing research mainly focuses on providing comfort through empathetic responses,rather than achieving specific real-world goals using these responses.To address this gap,this paper proposes a benchmark named SPEAKSMART,covering five scenarios to evaluate LLMs' ability to achieve real-world goals through highly empathetic responses in conversations.Subsequently,a two-dimensional evaluation framework based on provider satisfaction and requester satisfaction is introducted.Various LLMs are evaluated using SPEAKSMART and a baseline approach is designed to enhance their capabilities for generating empathetic and persuasive responses in conversations.Experiments reveal that Claude3 and LLaMA3-70B perform best across different scenarios,while other LLMs show room for improvement.This research lays the foundation for enhancing LLMs' ability to handle real-world tasks requiring highly empathetic responses to achieve goals.

Novel Discrete Diffusion Text Generation Model with Convex Loss Function

LI Sihui, CAI Guoyong, JIANG Hang, WEN Yimin

Computer Science. 2025, 52 (10): 231-238. doi:10.11896/jsjkx.240800147

Abstract

PDF(2056KB) ( 0 )

References | Related Articles | Metrics

Diffusion language models adopt a non-autoregressive generation approach that improves inference speed.Additionally,continuous refinement through an iterative refinement enhances the quality of the generated text,making it promising for text generation tasks.However,since diffusion language models are often trained using cross-entropy loss based on maximum likelihood estimation,even if the model generates a correct sentence,it may be penalized for not strictly aligning with the reference sentence,resulting in a serious multimodality problem,significantly reducing the quality of text generation.To alleviate the multimodality problem,a discrete diffusion language model ConvexDiffusion based on convex loss function training is proposed.The mo-del leverages the property of convex functions to sharpen the optimal distribution so that the model focuses more on high-probability outputs.To further improve the quality and reduce the repetition rate of generated words,a hybrid-aware noise schedule that enabled the noise labelling to vary non-linearly is designed,along with a high-confidence deterministic denoising strategy employed during the decoding process.Experimental results on the three text generation tasks－machine translation,question gene-ration,and question paraphrasing demonstrate that ConvexDiffusion achieves a performance improvement of 1~7 BLEU points and faster generation speed compared to leading diffusion models such as RDM and non-autoregressive models like CMLM.Especially on two large datasets,WMT16 EN-RO and WMT14 EN-DE,ConvexDiffusion surpasses the leading autoregressive models in text generation.

Multi-grained Sentiment Analysis of Comments Based on Text Generation

ZHANG Jiawei, WANG Zhongqing, CHEN Jiali

Computer Science. 2025, 52 (10): 239-246. doi:10.11896/jsjkx.240800025

Abstract

PDF(2726KB) ( 1 )

References | Related Articles | Metrics

With the rise of social media and online review platforms,automated sentiment analysis has become a key tool for understanding public emotions,consumer preferences,and market trends.Traditional sentiment analysis methods often use classification models that focus on extracting the overall sentiment of the text,neglecting the complex and multidimensional emotional information that may be contained within the comments.Addressing this issue,this study proposes a multi-granularity text-based sentiment analysis model using generative models to intricately capture aspect-level and document-level emotions in review texts.Additionally,a structured output format is constructed that includes sentiment labels for different aspects of the review text as well as the overall sentiment label of the review text.Compared to traditional classification models,the proposed model more comprehensively understands and reflects the emotional structure of text,achieving extraction and classification of multifaceted emotional information and overall sentiment in comments.Experimental results show that the proposed modelis better than conventional classification methods in the recognition of overall emotions and aspect emotions,and achieves a 4.4% higher F1-Score than the Bert+LSTM model.

Text Sentiment Classification Method Based on Large-batch Adversarial Strategy and EnhancedFeature Extraction

CHEN Jiahao, DUAN Liguo, CHANG Xuanwei, LI Aiping, CUI Juanjuan, HAO Yuanbin

Computer Science. 2025, 52 (10): 247-257. doi:10.11896/jsjkx.240800061

Abstract

PDF(2338KB) ( 0 )

References | Related Articles | Metrics

The text sentiment classification task aims to analyze short text sentences and determine their corresponding sentiment categories.In order to solve the problems of lack of large-scale high-quality corpus dataset and insufficient non-uniform importance extraction of text features in the existing models in sentiment classification,this paper proposes a text sentiment classification method based on large-batch adversarial strategy and enhanced feature extraction.Firstly,the text dataset is input into the pre-trained language model BERT to obtain the corresponding word embedding vector representation,and then the BiLSTM is used to further learn the context dependencies in the sequence.Then,the local attention mechanism is combined with the local receptive field weighting of TextCNN to enhance the feature extraction ability.Finally,the output of BiLSTM and the output of TextCNN are spliced to obtain the deep feature fusion of the two spaces,which are handed over to the classifier for the judgment of sentiment classification.In the whole training process,a large-batch adversarial strategy is adopted,and adversarial perturbations are added to the word embedding space and multiple iterations are carried out to improve the robustness of the model.Experimental results on multiple datasets verify the effectiveness of the proposed model.

Text Simplification for Aspect-based Sentiment Analysis Based on Large Language Model

WANG Ye, WANG Zhongqing

Computer Science. 2025, 52 (10): 258-265. doi:10.11896/jsjkx.250100114

Abstract

PDF(1993KB) ( 2 )

References | Related Articles | Metrics

Aspect-based sentiment analysis aims to identify the sentiment polarity of each aspect in a sentence.However,most existing approaches overlook the redundant and irrelevant information often present in review texts,which not only complicates model processing,but also hinders accurate sentiment element extraction.To address this issue,this paper proposes a model that transforms the original text into simplified clauses,expressing the same sentiment in a more concise manner.The key idea is to leverage a large language model to pre-identify aspect and opinion terms in the text,and then generate simplified clauses based on these identified sentiment elements.A self-verification mechanism is employed to ensure the generated clause satisfy three criteria:sentiment consistency,relevance,and conciseness.Furthermore,the model jointly uses both the original text and the simplified clauses to generate sentiment elements.Experimental results on public datasets—Restaurant,Laptop,and Phone,demonstrate that the model outperforms existing baselines,highlighting the significance of simplified clauses in aspect-based sentiment analysis.

Summary Faithfulness Evaluation Based on Data Augmentation and Two-stage Training

ZHAO Jinshuang, HUANG Degen

Computer Science. 2025, 52 (10): 266-274. doi:10.11896/jsjkx.250100023

Abstract

PDF(2136KB) ( 0 )

References | Related Articles | Metrics

The faithfulness of text summaries,which refers to their factual consistency with the original texts,is very important for the practical application of automatic text summarization.Current methods for evaluating the faithfulness of summaries have shortcomings in utilizing text summarization datasets,and the constructed unfaithful summaries differ significantly from the original texts,which limit the effectiveness of these evaluation methods.To solve this problem,this paper proposes a summary faithfulness evaluation model,FaithEval,based on data augmentation and two-stage training.Firstly,two data augmentation methods are defined:Similarity Search with Same Topic and Insert and Fill External Mask,which are used to generate summaries that are related but not faithful to the original texts.These methods are used to extract training data from the text summarization dataset.Secondly,to fully utilize the dataset information,the model is trained in two stages based on the training data constructed from the original texts and the reference summaries,progressively strengthening the faithfulness evaluation ability of the model.Finally,the test set for summary faithfulness evaluation SFETS,is constructed manually to provide a benchmark for testing model performance.Experiments show that FaithEval performs well on both SFETS and Rank19 datasets,and achieves the current state-of-the-art performance on the SFETS dataset.

GCE3S:A Method for Generating Safety-critical Scenarios in Autonomous Driving Based on Evolutionary Search

SUN Lele, HUANG Song, ZHENG Changyou, XIA Chunyan, YANG Zhen

Computer Science. 2025, 52 (10): 275-286. doi:10.11896/jsjkx.240800030

Abstract

PDF(3718KB) ( 2 )

References | Related Articles | Metrics

The rapid development of automated driving technology has brought great potential for transforming mobility,but automated driving technology,as safety-critical software,will lead to huge losses due to safety violations of self-driving vehicles in real traffic environments.In order to ensure that autonomous driving systems can operate safely in various complex traffic environments,autonomous driving systems must be fully tested before being deployed on real roads.Due to the complexity and high dimensionality of the autonomous driving test scenario space,existing safety critical scenario generation methods suffer from high cost and low efficiency.Therefore,this paper proposes an evolutionary search-based safety-critical scenario generation method for autonomous driving－GCE3S.GCE3S constructs safety-critical scenarios with adversarial nature by mapping the obstacles and their attributes in the scenario to the chromosome structure of genetic composition,thus perturbing the obstacles(vehicles,weather,pedestrians,etc.) in a more detailed manner and guiding the evolutionary search algorithms through multiple objective functions to generate diverse safety critical scenarios.In addition,the GCE3S is experimentally evaluated in the simulated environments of Baidu Apollo,an industrial-grade autonomous driving system,and LGSVL.The experimental results show that the number of safety-critical scenarios generated by GCE3S improve by 20.4% and the generated safety-critical scenarios increase by 20% in terms of diversity in the same amount of time as compared to the best baseline MOSAT method.

Multi-agent Formation Control Based on Discrete Layers of Formation Shapes

PAN Yunwei, LI Min, ZENG Xiangguang, XING Lijing, HUANG Ao

Computer Science. 2025, 52 (10): 287-295. doi:10.11896/jsjkx.240700193

Abstract

PDF(4227KB) ( 0 )

References | Related Articles | Metrics

A multi-agent system with multiple agents capable of completing complex tasks.In view of the formation of multi-agent in complex environment and the formation reorganization when the formation is impacted,a distributed formation control method based on the discrete layer of formation shape is proposed.Firstly,the formation shape is discretized and iterative,its influence range is expanded,and the formation information is shared with each agent.Secondly,for environments with obstacles,a dynamic negotiation algorithm is designed to adjust the formation's assembly position in real time.Finally,a speed controller is designed using sensor information and formation shape data,employing a distributed control method to achieve dynamic obstacle avoidance and manage complex formations.Experimental results show that the proposed method effectively guides multiple agents in forming complex formation shapes and enables formation obstacle avoidance,offset adjustment,and reorganization in environments with obstacles.Evaluation and analysis of the experimental results,using metrics for formation shaping time and perfor-mance,validate the method's strong environmental adaptability and effectiveness.

Sub-problem Effectiveness Guided Multi-objective Evolution Algorithm

SUN Liangxu, LI Linlin, LIU Guoli

Computer Science. 2025, 52 (10): 296-307. doi:10.11896/jsjkx.241000025

Abstract

PDF(3195KB) ( 0 )

References | Related Articles | Metrics

In order to solve the problems of poor performance and low universality in solving MOPs with unconventional PF based on decomposed multi-objective evolutionary algorithms,a new Multi-Objective Evolution Algorithm based on Decomposition and Sub-problem Effectiveness Guidance(MOEA/D-SEG) is proposed.The algorithm expands the sub-problem structure and describes the behaviour of weight vectors in the evolutionary process.The weight vector adjustment is realized by fission “efficient” sub-problems,so that the algorithm can better adapt to the multi-objective optimization problems with different characte-ristics,ensure the convergence and diversity of the solution set,and improve the ability of the algorithm to solve various complex multi-objective optimization problems.Through a series of experiments,the effectiveness of the proposed algorithm in different feature testing problems is proved,and the superiority of the proposed algorithm is proved by comparative analysis with other advanced algorithms.The application of the proposed algorithm in steel-making and continuous casting scheduling problem further verifies the feasibility.

Adaptive LQR Intelligent Vehicle Path Tracking Control Method Considering Time-varyingParameters

ZHANG Yajuan, FENG Lingxia, LI Guobin

Computer Science. 2025, 52 (10): 308-316. doi:10.11896/jsjkx.240800112

Abstract

PDF(4024KB) ( 0 )

References | Related Articles | Metrics

To solve the problem of vehicle tracking accuracy and stability degradation caused by model uncertainty resulting from tire cornering stiffness perturbation in the tracking process of intelligent vehicles,this paper proposes a tracking control method for intelligent vehicles considering the time-varying characteristics of tire cornering stiffness.Firstly,the tire lateral force estimator is established based on the improved set membership filter algorithm and the two-track dynamic model.The adaptive update rule of tire cornering stiffness is designed by using the difference between the calculated value and the estimated value of the tire lateral force model.Secondly,the real-time updated tire cornering stiffness is used to solve the real-time optimal front wheel angle,and then an ALQR intelligent vehicle tracking controller with adaptive tire cornering stiffness is proposed.The results of CarSim and Simulink co-simulation and hardware-in-the-loop simulation show that the proposed ALQR controller improves the tracking accuracy by 65.867% on average compared with LQR under the condition of high and low road adhesion coefficient.In particular,the LQR controller ignores the significant decline in tracking performance caused by changes in tire stiffness on low-adhesion roads.The proposed ALQR controller can solve the optimal front wheel angle in real time through real-time updated tire cornering stiffness to ensure vehicle tracking accuracy and stability.The tracking ALQR control method of intelligent vehicle considering real-time tire cornering stiffness proposed in this paper has good applicability.

WiLCount:A Lightweight Crowd Counting Model for Wireless Perception Scenarios

DUAN Pengsong, ZHANG Yihang, FANG Tao, CAO Yangjie, WANG Chao

Computer Science. 2025, 52 (10): 317-327. doi:10.11896/jsjkx.240800060

Abstract

PDF(4870KB) ( 1 )

References | Related Articles | Metrics

To address the challenges of limited accuracy and high computational complexity in crowd counting models due to the absence of spatial features in CSI,this paper proposes a lightweight model,WiLCount,based on amplitude-phase fusion.Firstly,a linear transformation method is applied to calibrate the phase data,addressing the issues of carrier frequency offset and sampling frequency offset in the raw phase information,which would otherwise render it unusable.Next,the amplitude-phase data is reconstructed into a two-dimensional image to fully exploit the spatial mapping features of crowd count inherent in CSI data.Finally,WiLCount is developed by integrating depthwise separable convolutions with a multi-branch structure.Due to the lack of publicly available datasets in the Wi-Fi-based crowd counting field,a self-collected dataset,leading the industry in terms of crowd scale and activity diversity,is meticulously constructed and released.Experimental results demonstrate that WiLCount achieves a recognition accuracy of up to 99.58% on the self-collected dataset,with a parameter size of only 4% of that of comparable mo-dels.Significant improvements over existing methods have been observed,with the model exhibiting strong robustness.

SRv6 Functional Conformance Verification Mechanism Based on the Programmable Data Plane

WANG Pengrui, HU Yuxiang, CUI Pengshuai, DONG Yongji, XIA Jiqiang

Computer Science. 2025, 52 (10): 328-335. doi:10.11896/jsjkx.240800163

Abstract

PDF(2399KB) ( 0 )

References | Related Articles | Metrics

At present,the SID in SRv6 is designed to provide programmability for traffic engineering,security authentication,and other network functions.The realization of these functions depends on the precise matching and execution of flow tables in the data plane,but when the flow tables are maliciously modified or incorrectly configured,it is easy to cause inconsistency problems in function implementation.As a classic verification tool with programmability in SDN scenarios,the INT technology can naturally combine with the two.This paper proposes the SRv6 Function Consistency Verification(SRv6FCV) mechanism based on programmable data plane.SRv6FCV uses data plane programmability technology to insert authentication identifiers into probe pac-kets,first dynamically converts the SID into a specific INT metadata structure according to the monitoring needs,then constructs probe packets and injects them into the network to collect flow table execution information for specific network functions,and finally decodes the telemetry information and completes the function consistency verification based on symbolic execution algorithms.Simulation results show that SRv6FCV can ensure consistency between flow table rules and business function execution policies.Compared with previous studies,SRv6FCV,in addition to achieving consistency verification of network functions,has lower running overhead and significantly reduces verification time.

Wireless Charging Scheduling with Minimized Maximum Return-to-Work Time for Heterogeneous Mobile Rechargeable Devices

XU Jia, LIU Jingyi, XU Lijie, LIU Linfeng

Computer Science. 2025, 52 (10): 336-347. doi:10.11896/jsjkx.240800113

Abstract

PDF(2978KB) ( 2 )

References | Related Articles | Metrics

With the widespread application of wireless rechargeable devices,wireless power transfer technology has become a key enabler for enhancing device battery life.However,most existing research focuses on optimizing charging efficiency,charging costs,and other charging-related performance metrics.Task-driven charging scheduling has received limited attention.In addition,the majority of charging systems assume homogeneity either in chargers or rechargeable devices,and there is a lack of focus on charging scheduling for mobile rechargeable devices.Considering the heterogeneity of rechargeable devices in real-world scena-rios,this paper proposes a heterogeneous wireless charging model and formalizes the problem of minimizing the maximum return-to-work time of mobile rechargeable devices in heterogeneous wireless rechargeable sensor networks.Initially,it studies the simplified problem of minimizing the maximum return-to-work time under the assumption of ignoring device movement time and energy consumption,and proposes an approximation algorithm.To address the more challenging problem of minimizing the maximum return-to-work time,this paper proposes a heuristic algorithm based on the idea of this approximate algorithm for this problem.The results of extensive simulation and field experiments demonstrate that the proposed algorithm has significant advantages over other algorithms.The proposed algorithm can reduce at most 38.78% maximum return-to-work time of devices comparing with the benchmark algorithms.

Approach for Lightweight Verifiable Data Management Based on Blockchains

WU Moxun, PENG Zeshun, YU Minghe, LI Xiaohua, DONG Xiaomei, NIE Tiezheng, YU Ge

Computer Science. 2025, 52 (10): 348-356. doi:10.11896/jsjkx.250200001

Abstract

PDF(2175KB) ( 3 )

References | Related Articles | Metrics

Combining the advantages of both blockchain and database technologies,this paper proposes a blockchain-based verifiable data management approach.Unlike existing works,this approach does not require changes to the existing database storage mode,allowing for verifiable queries and data traceability without changing the performance and structure of the original database,thus better suiting various application scenarios.Firstly,a two-phase commit protocol is adopted to design an interaction model among users,the database,and the blockchain,along with an efficient data synchronization mechanism to ensure the atomicity of transactions between the blockchain and the database.Secondly,an index tree structure based on data attribute encoding is introduced to support accurate,complete,and efficient verifiable queries on the blockchain.Finally,comprehensive theoretical analysis and performance evaluations are conducted.In terms of security,it achieves data tamper-resistance and user privacy protection.In terms of system effectiveness,it ensures consistency between blockchain and database data.In terms of system efficiency,it reduces storage space by about 90% and improves query verification efficiency by approximately 77% compared to other typical works.

Intrusion Detection Method Based on Improved Active Learning

HE Hao, ZHANG Hui

Computer Science. 2025, 52 (10): 357-365. doi:10.11896/jsjkx.240900142

Abstract

PDF(2495KB) ( 0 )

References | Related Articles | Metrics

Conventional intrusion detection methodologies based on deep learning necessitate a substantial number of labeled samples to attain optimal accuracy.Nevertheless,the acquisition of a substantial number of labeled samples necessitates a considerable investment of time and labor,which constrains its applicability in practical settings.In order to address these limitations,a novel intrusion detection method that integrates active learning with convolutional neural networks is proposed.This method employs an enhanced adaptive active learning approach to more efficiently identify the most representative samples for labeling,effectively reducing the computational cost of the model training process and enhancing the overall performance of the model.The experimental results on the CCF-BDCI-2022 and Malicious-URLs-2021 datasets demonstrate that the proposed method exhibits superior performance in terms of query time and iteration time compared to traditional deep learning-based models.In the CCF-BDCI-2022 dataset,the method demonstrates an accuracy rate of 97.10% and a false positive rate of 1.3%.In the Malicious-URLs-2021 dataset,the method achieves an accuracy rate of 99.05% and a false positive rate of 1.1%.Compared with other methods,this method not only performs better in terms of accuracy and false positive rate,but also significantly reduces the consumption of computing resources,thereby improving the efficiency and practicality of the model.

Boundary Black-box Adversarial Example Generation Algorithm on Video Recognition Models

JING Yulin, WU Lijun, LI Zhiyuan, DENG Qi

Computer Science. 2025, 52 (10): 366-373. doi:10.11896/jsjkx.240700045

Abstract

PDF(2199KB) ( 0 )

References | Related Articles | Metrics

With the rapid development of deep learning,neural networks are widely used in various fields.However,neural networks still face the problem of adversarial attacks.Among all types of adversarial attacks,the boundary black-box attack can only obtain the final classification label of the tested model,so it is closest to the actual application scenario,and is recognized as the most practical and difficult attacks,which has attracted more and more researchers to conduct related research.Nevertheless,current relevant research mainly focus on image recognition models,and with less research on video recognition models.To this end,this paper proposes a boundary black-box video adversarial example generation algorithm BBVA.BBVA uses a progressive exploration mechanism to generate adversarial videos,which effectively improves the efficiency of updating samples.Experiments show that compared with the state-of-the-art boundary black-box video adversarial example generation algorithm STDE,BBVA better balances the noise size and model queries,and gets the best results in this research field in many measurement indicators such as visual effect,optimization distance and fooling rate.In addition,under more severe conditions,BBVA even outperforms some state-of-the-art score-based black-box video adversarial example generation algorithms,such as EARL and VBAD.The proposed algorithm can be used to provide adversarial training samples to enhance video model security.

High-frequency Feature Masking-based Adversarial Attack Algorithm

WANG Liuyi, ZHOU Chun, ZENG Wenqiang, HE Xingxing, MENG Hua

Computer Science. 2025, 52 (10): 374-381. doi:10.11896/jsjkx.241000030

Abstract

PDF(2758KB) ( 1 )

References | Related Articles | Metrics

Deep neural networks have achieved widespread application in the field of imagerecognition,however,their complex structures make them vulnerable to adversarial attacks.Constructing adversarial examples that are imperceptible to the human eye is crucial for evaluating the security of these networks.Existing adversarial example generation methods for images typically involve small perturbations to the original samples,constrained by l_p-norms.This simplistic approach treats all pixels equally,applying uniform constraints to the allowable perturbations at each pixel,which limits the flexibility of adversarial example generation and makes the perturbations more detectable to the human eye.In practical applications,human visual sensitivity varies across different colors and image regions.To address this issue,this paper proposes an adaptive perturbation scheme based on perceptual sensitivity,where different perturbation constraints are applied to different pixels,thereby enhancing the robustness of the adversarial examples.Specifically,this method employs spectral analysis to divide the image into high-frequency and low-frequency regions and applies novel spatial constraints to regulate perturbations.Larger perturbations are introduced in regions less sensitive to high-frequency changes,improving adversarial effectiveness.Extensive experiments conducted on the ImageNet-1K and CIFAR-10 datasets demonstrate that the proposed adversarial example generation strategy can be coupled with various attack me-thods,significantly enhancing adversarial performance while ensuring imperceptibility.

Benign-salient Region Based End-to-End Adversarial Malware Generation Method

YUAN Mengjiao, LU Tianliang, HUANG Wanxin, HE Houhan

Computer Science. 2025, 52 (10): 382-394. doi:10.11896/jsjkx.240800046

Abstract

PDF(4657KB) ( 2 )

References | Related Articles | Metrics

Malware detection methods combining visualization techniques and deep learning have gained widespread attention due to their high accuracy and low cost.However,deep learning models are vulnerable to adversarial attacks,where intentional small-scale perturbations can misguide the model into making incorrect decisions with high confidence.Current research on adversarial attacks targeting visualization-based detection methods for Windows malware has primarily focused on improving the effectiveness of adversarial images,while neglecting the actual harmfulness of adversarial examples.Therefore,this study proposes a novel method to generate harmful adversarial malware,BREAM(Benign-salient Region based End-to-end Adversarial Malware generation method).Firstly,selecting the salient regions of benign images as initial perturbations to enhance the attack effect on adversarial images,and introducing a mask matrix to restrict the perturbation range to ensure the functionality of adversarial examples.Then,an inverse feature mapping method is proposed to convert adversarial images into adversarial malware,achieving end-to-end generation of malware adversarial examples.The attack performance of BREAM is evaluated on four target models,and experimental results show that when the target models employ bilinear interpolation and nearest neighbor interpolation respectively,compared with existing methods,the attack success rate of adversarial images generated by BREAM has increased by an average of 47.96% and 28.39%;the attack success rate of adversarial malware has increased by an average of 53.25% and 61.93%,causing the classification accuracy of the target models to decrease by an average of 92.82% and 73.64%.

Dual-stream Feature Fusion Approach for Dockerfile Security Misconfiguration Detection

ZHAO Ning, WANG Jinshuang, CUI Shuai

Computer Science. 2025, 52 (10): 395-403. doi:10.11896/jsjkx.241000014

Abstract

PDF(2467KB) ( 0 )

References | Related Articles | Metrics

Dockerfile misconfigurations frequently lead to container security vulnerabilities.Current detection methods rely on structural analysis and semantic understanding of the text,while pay little attention to metrics such as command frequency,image layer counts,code complexity,etc.To solve this problem,a dual-stream deep learning detection approach is proposed,which integrates feature metrics with semantic comprehension.Firstly,it identifies and annotates Dockerfile samples containing security misconfigurations using static detection tools such as Hadolint and KICS.Then,by constructing abstract syntax trees,it parses and extracts code metric features and refines crucial security features using the random forest algorithm.Lastly,it extracts textual information and security feature metrics and then inputs them into a dual-stream model for detection.Bi-LSTM network is utilized to trace the forward and backward dependencies within instruction sequences,which is helpful for uncovering deep semantic associations.Transformer model is employed to create high-dimensional metric representations,which can model mappings from me-tric to security misconfiguration.CNN sublayers with ReLU activation functions are used to fuse information from both streams.Experimental results indicate that the proposed method achieves 96%,98% and 97% in precision,recall,and F1-score respectively.The proposed approach can detect security misconfiguration more accurately compared to existing approaches.

Security-aware Service Function Chain Deployment Method Based on Deep ReinforcementLearning

ZHU Ziyi, ZHANG Jianhui, ZENG Junjieand ZHANG Hongyuan

Computer Science. 2025, 52 (10): 404-411. doi:10.11896/jsjkx.240800015

Abstract

PDF(3304KB) ( 2 )

References | Related Articles | Metrics

As a key technology to improve the utilization of network resources,service function chain combined with deep reinforcement learning makes it possible to achieve flexible and secure deployment.However,how to effectively deploy service function chains with security requirements while maximizing long-term average revenue is an important challenge it faces.This paper proposes a deployment method for security-aware service function chain based on deep reinforcement learning(DRL-SASFCD).Firstly,a security-aware mechanism is proposed to evaluate the credibility of physical network nodes,and a security requirement index is introduced to perceive the security requirements of SFC.Secondly,this method utilizes graph attention network and sequence to sequence models to extract relevant features of underlying physical network information and service function chain request sequence information.It generates service function chain deployment strategies based on these features.Finally,the proximal policy optimization method is adopted to optimize the policy and training network parameters.By limiting the update amplitude between the new and old policies,the drastic fluctuations during the policy update process are avoided,thereby improving the efficiency of security policy optimization.The simulation results show that DRL-SASFCD can improve the deployment acceptance rate,long-term average revenue and long-term average revenue-cost ratio compared with the existing methods while considering the security requirements of service function chain deployment.

Multi-functional Attribute Based Encryption from Lattices

GUO Lifeng, YANG Jieying, MA Tianjun, ZHANG Xialei

Computer Science. 2025, 52 (10): 412-422. doi:10.11896/jsjkx.240600137

Abstract

PDF(1900KB) ( 1 )

References | Related Articles | Metrics

Attribute based encryption from lattices has the property of resisting quantum attacks,and achieves fine-grained access control of attributes by cleverly embedding access control policies into ciphertext or keys.However,due to the inherent weaknesses of attribute based encryption,users with the same attribute may leak the key.To avoid key leakage,attribute based encryption schemes need to implement the function of tracking and revoking specific user decryption permissions.However,illegal users may still attempt to recover the keys of past sessions by collecting large amounts of encrypted data.To effectively resist such attacks,the scheme must implement forward security.In response to the current demands and challenges in the field of lattices cryptography,this paper proposes a multi-functional attribute based encryption scheme based on the Decisional Learning with Errors(DLWE) problem from lattices that can prove security.The scheme uses a complete binary tree to track the identity matrices related to the users in the decryption key(such as the values of the leaf nodes of the complete binary tree) in order to track malicious users.Introducing a user revocation mechanism that allows attribute authority to revoke user permissions in a timely and effective manner without generating new keys for the users.Using tag puncturing method to ensure that even if the current key is leaked,the past ciphertext remains secure and achieves forward security.In addition,due to the uncertainty of the upsampling algorithm from lattice,it is currently difficult to achieve experiments on attribute based encryption from lattice.Therefore,the security and correctness of the scheme are verified through theoretical analysis.The scheme not only optimizes space storage efficiency,but also compensates for the shortcomings caused by the lack of functions of attribute based encryption schemes on lattice cryptography.

DLSF:A Textual Adversarial Attack Method Based on Dual-level Semantic Filtering

XIONG Xi, DING Guangzheng, WANG Juan, ZHANG Shuai

Computer Science. 2025, 52 (10): 423-432. doi:10.11896/jsjkx.240700202

Abstract

PDF(2387KB) ( 1 )

References | Related Articles | Metrics

In the field of commercial applications,deep learning-based text models play a crucial role but are also susceptible to adversarial samples,such as the incorporation of confusing vocabulary into reviews leading to erroneous model responses.A strong attack algorithm can assess the robustness of such models and test the effectiveness of existing defense methods,thereby reducing potential harms from adversarial samples.Considering the prevalent issues of low-quality adversarial texts and inefficient attack methods in black-box settings,this paper proposes a dual-level semantic filtering attack algorithm based on word substitution.This algorithm amalgamates existing methodologies for assembling candidate word sets,effectively eliminates interfe-rence from irrelevant words,and thereby enriches the variety and quantity of candidate words.It employs a dual-filter beam search strategy during the iterative search process,which not only reduces the frequency of model access,but also guarantees the acquisition of optimal adversarial texts.Experimental results on text classification and natural language inference tasks demonstrate that this method significantly enhances the quality of adversarial texts and attack efficiency.Specifically,the attack success rate on the IMDB dataset reaches 99.7%,semantic similarity reaches 0.975,with the number of model accesses being only 17% of those required by TAMPERS.Furthermore,after adversarial augmentation training with adversarial samples,the target model's attack success rate on the MR dataset decreases from 92.9% to 65.4%,further confirming that DLSF effectively enhances the robustness of the target model.