Computer Science

Survey of Backdoor Attacks and Defenses on Graph Neural Network

DING Yan, DING Hongfa, YU Muran, JIANG Heling

Computer Science. 2026, 53 (3): 1-22. doi:10.11896/jsjkx.250700093

Abstract

PDF(5345KB) ( 336 )

References | Related Articles | Metrics

In artificial intelligence(AI)-driven intelligent information systems,GNNs are extensively utilized for knowledge discovery and decision support in critical domains including social network analysis and financial risk control,leveraging their superior graph-structured data modeling capabilities.However,the heavy reliance of such systems on third-party data and models exposes GNNs to stealthy backdoor attacks.Attackers can inject backdoor triggers or tamper with models to induce predetermined erroneous outputs for inputs containing specific patterns,thereby undermining the trustworthiness and reliability of intelligent information services.To ensure the security and controllability of intelligent information systems,this paper systematically reviews research on GNN backdoor attacks and defenses through dual data-model perspectives.It firstly conducts in-depth analysis of attack vectors during data collection,model training,and deployment phases,establishing a comprehensive attack-defense framework.It subsequently categorizes attacks into six data-oriented and two model-oriented types based on implementation stages and attacker capabilities,classifies defenses into data-oriented,model-oriented,and robust training-oriented approaches according to deployment stages and defender capacities,with detailed comparative examination of their core mechanisms,technical features,advantages,and limitations.Finally,it summarizes current research challenges while outlining future directions.The proposed attack-defense taxonomy facilitates profound understanding of GNN backdoor threat evolution and advances security design for next-generation trustworthy intelligent information systems.

Optimization of Service Level Objectives and System Level Metrics in Large Language ModelServing System

WANG Zhibin, LI Shipeng, ZHOU Yuhang, LI Xue, ZHANG Zhonghui, JIANG Zhiwei, GU Rong, TIAN Chen, CHEN Guihai, ZHONG Sheng

Computer Science. 2026, 53 (3): 23-32. doi:10.11896/jsjkx.250900173

Abstract

PDF(3130KB) ( 203 )

References | Related Articles | Metrics

In Large Language Model(LLM) serving systems,user experience is a critical consideration.Service-Level Objectives(SLOs) and System-Level Metrics(SLMs) are two key performance measures:the former focuses on the experience of individual requests,while the latter reflects the overall performance of the system.However,existing metrics exhibit two counterintuitive issues:1) manually delaying the delivery of some tokens can improve SLOs;2) actively abandoning requests that do not meet SLOs can improve SLMs.To address these issues,the definitions of SLOs and SLMs in LLM serving are revisited and a new type of SLO is proposed that aligns more closely with actual user experience.Based on this SLO,a comprehensive metric framework called smooth goodput is developed,which integrates SLOs and SLMs to reflect the nature of user experience in LLM serving.Through this unified framework,the performance of different LLM serving systems under multiple workloads is reassessed.Eva-luation results show that the proposed metric framework provides a more comprehensive view of token delivery and request processing,and effectively captures the optimal point of user experience and system performance with different serving strategies.

Training System for Large Language Models Based on Adaptive Transpose on Hygon DCU

ZHOU Yueyuan, LU Guanze, XIANG Jiawei, ZHANG Jiawei, SHAO En, HE Xin

Computer Science. 2026, 53 (3): 33-40. doi:10.11896/jsjkx.250600073

Abstract

PDF(4132KB) ( 188 )

References | Related Articles | Metrics

With the intensification of trade frictions between China and the United States,the development of domestic accelerator chips in China has become increasingly urgent.The Hygon DCU,with its CUDA-like architecture,excellent compatibility,and cost-effectiveness,has emerged as a strong candidate to replace high-end American chips in the field of artificial intelligence.However,on the Hygon DCU platform,the performance of the GEMM kernel function,which is a critical operator in large language model training,varies significantly.This paper investigates the impact of matrix transposition on the performance of the GEMM kernel function in the rocBLAS algorithm library and proposes two optimization methods:minimizing transposition and adaptive transposition,to effectively reduce the training time of large language models.This study modifies the implementation of the linear layer in PyTorch and proposes the minimization and adaptation of transposition methods for distributed training of large language models.Experimental results show that these two optimization methods significantly reduce training time in the distri-buted training of various large-scale language models,such as OPT-6.7B,LLaMA-7B,and Bloom-7B.Among the 83 test cases,the adaptive transposition method outperformes in 72 cases,with the highest improvement of 24.27% in end-to-end training time compared to the original PyTorch-based Megatron-LM.

Large Language Model and Deep Network Based Cognitive Assessment Automatic Diagnosis

CHEN Han, XU Zefeng, JIANG Jiu, FAN Fan, ZHANG Junjian, HE Chu, WANG Wenwei

Computer Science. 2026, 53 (3): 41-51. doi:10.11896/jsjkx.250600034

Abstract

PDF(4674KB) ( 177 )

References | Related Articles | Metrics

Cognitive assessment is one of the important assessment tools for rapid screening of cognitive impairment.Traditional method relies on the experience and judgment of doctors,which is difficult to ensure the objectivity and accuracy of diagnosis results.The development of deep network technology and the rise of large language model have promoted the progress of medical intelligent auxiliary diagnosis.It is of great significance to carry out the research on automatic auxiliary diagnosis of medical cognitive assessment.Aiming at this issue,this paper focuses on a commonly used cognitive assessment Montreal Cognitive Assessment(MoCA),and proposes a framework consisting of large language model and deep network based image classification model for automatic diagnosis of MoCA and selects base model under this framework.To improve the processing abilities of the base models for the evaluation of assessment questions,this paper proposes CSWin-FLA Transformer(Cross-Shaped Window with Focused Linear Attention Transfromer) and AGPoFS(automatic generation of prompts based on fewer samples),and designs a MoCA diagnosis process.Since there is no public MoCA dataset,the assessment data provided by Zhongnan Hospital of Wuhan University are collected to form the datasets.Experiments are carried out from each method to the overall system,and the best application performance is achieved on the proposed datasets,which proves the effectiveness of the relevant improvements and verifies the effectiveness of the system.

SQL-MARS:Text-to-SQL Structured Data Recommendation System for Ambiguous UserRequirements

XU Jiawen, ZHENG Yungui, ZHOU Wei, XU Yaoqiang, HU Huiqi, ZHOU Xuan

Computer Science. 2026, 53 (3): 52-63. doi:10.11896/jsjkx.250700096

Abstract

PDF(2845KB) ( 197 )

References | Related Articles | Metrics

With the maturity of LLM technology,natural language-based database interaction systems(e.g.,Chat2DB,ChatExcel) have achieved wide application.However,existing systems generally rely on the “precise query” assumption and struggle to handle the ubiquitous ambiguous requirements in real-world scenarios,where users need to clarify their query needs during interaction with the system.To address this challenge,this paper proposes SQL-MARS(SQL-oriented Multi-Agent Recommender System),a multi-agent collaborative framework based on a “perception-action-evaluation” closed-loop mechanism for dynamic identification and adaptive processing of ambiguous database query requirements.The system introduces a three-layer metadata architecture to model user’s requirements for ambiguous awareness.Based on this,it implements data navigation function,providing query recommendations at varying granularities based on users’ ambiguous requirements to progressively guide them in clari-fying their query needs.Additionally,the system proposes the fusion mechanism between external knowledge and local data to fully utilize valuable information from external sources.We alsocreate the dataset named Bird-fuzzy for ambiguous requirements and implements automated evaluation.Experimental results show that SQL-MARS can effectively identify ambiguous requirements and guide users to clarify their data needs.

Node-influence Based Construction Algorithm of Approximate Worst-case Forgetting Set for Graph Unlearning

ZHAO Zhengbiao, LU Hanyu, DING Hongfa

Computer Science. 2026, 53 (3): 64-77. doi:10.11896/jsjkx.250700094

Abstract

PDF(2171KB) ( 135 )

References | Related Articles | Metrics

GNNs have attracted significant attention due to their wide-ranging applications in fields such as social networks and recommendation systems.Motivated by intensified demands for data removal,including the right to be forgotten,data-ownership safeguards,and the expiration of data-use permissions,research on unlearning has accelerated,particularly in graph unlearning,deep unlearning,and unlearning in large language models.However,most existing studies simulate forgetting by randomly removing data,which fails to provide the strongest guarantee for data owners’ right to be forgotten and neglects the construction of extreme scenarios required to comprehensively evaluate different unlearning algorithms.To address these issues,this paper proposes a worst-case forgetting set construction algorithm based on node influence in graph data,aiming to approximately construct an optimal set of nodes to be forgotten for graph unlearning.This algorithm ranks the influence of training sample nodes by combining each node’s training loss and structural centrality,thereby identifying the most influential and hardest-to-forget nodes.It then selects the forgetting set by jointly considering each node’s impact on model performance and its importance.Experiments on different GNN models,graph datasets,and unlearning algorithms show that the proposed algorithm enables graph unlearning methods to more effectively reduce model performance.And it achieves up to a 15% greater reduction in performance compared to random forgetting strategies.At the same time,the approach significantly accentuates performance differences among various graph unlearning algorithms across multiple metrics,facilitating more effective multi-dimensional evaluation of these unlearning algorithms.

Fatigue Driving Detection Based on Dual-branch Fusion and Segmented Domain AdaptationTransfer Learning

LI Zequn, DING Fei

Computer Science. 2026, 53 (3): 78-87. doi:10.11896/jsjkx.250500025

Abstract

PDF(4566KB) ( 145 )

References | Related Articles | Metrics

Fatigue driving is one of the leading causes of traffic accidents.To address the issues of insufficient feature extraction caused by factors such as camera angles and environmental lighting,as well as poor model adaptability across different datasets,this paper proposes a novel transfer learning-based fatigue driving detection framework.The framework employs a dual-branch feature extraction and fusion architecture combining CNN and Transformer,leveraging their complementary strengths to enhance feature representation and comprehensively capture both local and global facial features of drivers.To improve the model’s adaptive capability between source and target domains,a segmented domain adapt-ation strategy is adopted.Adversarial domain adaptation and multi-kernel maximum mean discrepancy(MK-MMD) are applied during the feature extraction stage,while MK-MMD and minimum class confusion(MCC) loss are further introduced during the feature fusion stage to enhance cross-domain adapt-ability.Experimental results on two datasets with significant feature disparities demonstrate that the proposed framework achieves6 detection accuracies of 93.3%(A→B) and 75.1%(B→A) on the target domain,significantly improving the model’s adaptability and robustness.

Group Semantic-driven Hypergraph Network for Disinformation Detection with Fusion PropagationStructure

CUI Mengtian, HE Liwen, XIE Qi, WANG Fang

Computer Science. 2026, 53 (3): 88-96. doi:10.11896/jsjkx.250800013

Abstract

PDF(3805KB) ( 123 )

References | Related Articles | Metrics

In social networks characterized by frequent and intensive user interactions,disinformation tends to propagate rapidly through collaborative diffusion,exhibiting complex multi-level propagation structures and semantic associations.This represents one of the critical challenges urgently needing to be addressed in the field of national security technology.However,current detection methods,limited to either textual content or conventional propagation graphs,fail to capture these high-order semantic interactions and collaborative behaviors.Therefore,this paper proposes a group-semantics-driven hypergraph network method(GSHN-DD) that integrates propagation structures.The proposed method first constructs an initial hypergraph based on user behaviors and information topics to capture group-level coordination and semantic associations.Subsequently,latent higher-order hyperedges are mined through link prediction combined with a dual-layer filtering mechanism,resulting in an enhanced hypergraph topology.Building on this foundation,a hypergraph convolutional network,combined with a dual-layer attention mechanism,is utilized to integrate global group propagation patterns and local key hyperedge features.Finally,the model integrates propagation features and hypergraph semantic representations to generate unified embeddings,which are fed into a fully connected classifier for disinformation detection.Experimental results on the PolitiFact and GossipCop datasets demonstrate that GSHN-DD performs better than the baseline methods,achieving 2 to 5 percentage point improvement in accuracy and 2 to 7 percentage point increase in F1-score.

Joint Entity and Relation Extraction Method with Multi-scale Collaborative Aggregation and Axial-semantic Guidance

QIAN Qing, CHEN Huicheng, CUI Yunhe, TANG Ruixue, FU Jinmei

Computer Science. 2026, 53 (3): 97-106. doi:10.11896/jsjkx.250500095

Abstract

PDF(4098KB) ( 136 )

References | Related Articles | Metrics

In recent years,table-filling approaches to joint entity-relation extraction have achieved impressive performance,yet they typically neglect two critical challenges:modelling boundary correlations among token pairs and distinguishing semantically similar token pairs.To address these gaps,this paper introduces a novel joint extraction model featuring multi-scale semantic aggregation and axial-semantic guidance.Firstly,multi-scale semantic aggregation module applies parallel depthwise convolutions of varying kernel sizes to capture boundary correlation information across multiple spatial arrangements,thereby enriching token-pair representations and facilitating the detection of implicit entities.Next,axial-semantic guidance module employsrow-and co-lumn-wise banded convolutions to perform axis-aligned attention calibration,strengthening key semantic features and effectively resolving high-similarity ambiguities.Comprehensive experiments on NYT^*,WebNLG^*,NYT,and WebNLG datasets yield F1 scores of 93.2%,94.5%,93.2%,and 91.4%－corresponding to absolute gains of 0.1 percentage points,0.6 percentage points,0.4 percentage points,and 1.0 percentage points over strong baselines.These results validate that explicitly capturing multi-scale boundary correlations and refining semantic alignment substantially enhances joint entity-relation extraction.

Multi-view Exercise Representation and Forgetting Mechanism for Deep KnowledgeTracing

YU Chengcheng, JIANG Yongfa, CHEN Fangshu, WANG Jiahui, MENG Xiankai

Computer Science. 2026, 53 (3): 107-114. doi:10.11896/jsjkx.250700092

Abstract

PDF(2896KB) ( 124 )

References | Related Articles | Metrics

Knowledge tracing is a core task in intelligent tutoring systems,which aims to model a learner’s mastery of knowledges based on their historical interaction behaviors and predict the next exercise answer.However,existing approaches suffer from three main limitations:1)Most methods rely heavily on exercise or knowledge IDs,lacking the ability to capture complex structural dependencies between exercises and knowledge concepts to enhance exercise representation;2)They fail to effectively utilize multi-dimensional exercise attributes to enrich exercise embeddings;3)They overlook the impact of forgetting patterns in learners’ cognitive processes,which limits predictive performance.To address these issues,Multi-view Exercise representation and Forgetting mechanism for deep Knowledge Tracing(MEFKT)is proposed.Specifically,MEFKT leverages pre-trained models to learn high-quality exercise embeddings and predicts learner responses based on learning dynamics.Firstly,a graph-based contrastive learning strategy is adopted to pre-train exercise representations enriched with structural information.At the same time,it constructs attribute-enhanced pre-trained embeddings that capture multi-dimensional features such as exercise/knowledge simila-rity,exercise difficulty,exercise type,and response time.These multiple perspectives are then fused into a unified representation space via a learnable linear alignment module.Finally,a behavior prediction module incorporating a forgetting mechanism is designed to dynamically update knowledge states and predict the next response.Extensive experiments on two public benchmark datasets demonstrate that MEFKT significantly outperforms existing state-of-the-art knowledge tracing models,validating the effectiveness of integrating multi-view exercise representations and forgetting-aware learning dynamics.

Review of Methods and Applications of Graph Diffusion Models

ZHAO Haihua, TANG Rui, MO Xian

Computer Science. 2026, 53 (3): 115-128. doi:10.11896/jsjkx.250200118

Abstract

PDF(3305KB) ( 134 )

References | Related Articles | Metrics

Graph diffusion models,as an emerging paradigm in deep generative modeling,have demonstrated remarkable advantages in modeling complex graph-structured data due to their progressive generation mechanisms and structural flexibility.This paper systematically reviews the methodological evolution and application advancements of graph diffusion models.Firstly,three core paradigms are analyzed from the perspective of generative mechanisms:denoising diffusion probabilistic models,score-based diffusion generative models,and stochastic differential equation(SDE)-based diffusion generative models.Subsequently,to address the high-dimensional,discrete,and non-Euclidean nature of graph data,innovative technical breakthroughs of these three fundamental diffusion models in graph data processing are categorized,summarized,and subjected to in-depth analysis.Building on this foundation,the evaluation frameworks for graph diffusion models are systematically summarized and analyzed.At the application level,the study focuses on the applications of graph diffusion models in recommendation systems and molecular modeling.Finally,based on the above discussions,prospects for future challenges and potential research directions are proposed,encompassing four aspects:the discrete nature of graph data,conditional generation of graph diffusion models,application expansion,and evaluation frameworks.

Twice Learning Revitalizes Behavior Cloning

FAN Wenshu, WAN Shenghua, LI Xinchun, SUN Haihang, HUANG Kaichen, GAN Le, ZHAN Dechuan

Computer Science. 2026, 53 (3): 129-135. doi:10.11896/jsjkx.250600131

Abstract

PDF(3076KB) ( 110 )

References | Related Articles | Metrics

In the imitation learning method of behavior cloning(BC),an agent tends to take random actions when encountering states that are not covered by expert data.This deviation from the expert policy leads to what is known as compounding error,a critical factor affecting the performance of BC.To address this issue,this paper first establishes that BC can be regarded as a simplified form of twice learning.Furthermore,in discrete action environments,BC primarily focuses on aligning with the expert-selected actions while ignoring probability information associated with other actions,resulting in incomplete extraction of expert knowledge.Inspired by twice learning,this paper proposes an enhanced version of BC,termed complete behavior cloning(CBC),which aims to leverage a more comprehensive set of information from expert data.To validate the effectiveness of this approach,this paper designs multiple comparative experiments.The results demonstrate that CBC not only mitigates compounding error but also exhibits high transferability across different devices,enhanced robustness to noise,and reduced dependency on expert data.These findings suggest that BC can become highly practical and computationally efficient with only minor modifications.More-over,the experimental results further reinforce the guiding role and effectiveness of twice learning in reinforcement learning problems.

Data Compression of Instruction Fine-tuning for Large Models:Refinement Based on Inference Contribution

LI Hao, DING Lizhong, FU Jiarun, LINGHU Zhaohuan

Computer Science. 2026, 53 (3): 136-142. doi:10.11896/jsjkx.250600087

Abstract

PDF(2335KB) ( 97 )

References | Related Articles | Metrics

The fine-tuning of large model instructions based on reasoning data significantly improves the reasoning accuracy of the model by explicitly modeling the multi-step logical correlations of complex tasks.However,the fine-tuning process relies on massive high-quality data,resulting in a sharp increase in computing power overhead.The existing data compression techniques mainly focus on the reduction of the original scale.Generally,there is a lack of compression method design for reasoning data,ignoring multi-step logical associations,semantic dependency relationships in reasoning data,resulting in the damage of the integrity of the key reasoning chain and thereby reducing the reasoning performance.To this end,refinement based on inference contribution(RBIC) is proposed.The knowledge domain graph is constructed by analyzing and inferring the semantic similarity of the data to accurately locate the core information.It combines the semantics of data samples with the reasoning accuracy of large models,divides the difficulty gradient,and covers the reasoning requirements of all scenarios.The reasoning contribution is quantified through the logical complexity of multi-step reasoning data,and the data samples that contribute the most to the model’s reaso-ning are refined.Experimental results show that after fine-tuning based on the reasoning data refined by RBIC,the average reaso-ning performance of the model only decreases by 1.13%,while the training time is shortened to 16% of the original time consumption.This verifies that RBIC achieves the optimal balance between model efficiency and resource consumption,and is expected to promote the efficient deployment and fine-tuning optimization of multi-domain large models in resource-constrained environments.

Bayesian Network Based Fault Root Cause Analysis

LIU Huashuai, TAO Houguo, YUE Kun, DUAN Liang

Computer Science. 2026, 53 (3): 143-150. doi:10.11896/jsjkx.241200100

Abstract

PDF(2134KB) ( 98 )

References | Related Articles | Metrics

Fault root cause analysis is to find the occurrence cause of specific problems,faults and events,becoming the important technique for origin tracing in several paradigms.However,existing methods still cannot satisfy practical requirements of efficiency,accuracy and stability.BN(Bayesian network) is used as the knowledge framework for representing and inferring the depen-dencies among relevant attributes.Specifically,the vector quantized variational autoencoder algorithm for attribute reduction is proposed at first.Then,the α-BIC scoring metric is adopted to learn RCBN efficiently.Following,efficient inferences in RCBN are implemented by BN embedding by calculating the probabilities of fault occurrence for given causes.Finally,the Blame mechanism in causal model is adopted to evaluate the contribution of causes w.r.t.given faults and fulfill fault root cause analysis.Experimental results on 3 public datasets and 3 synthetic datasets show that the average accuracy and efficiency of the proposed fault detection are better than current representative methods,such that the precision is 7% higher and the running time is 60% faster than the comparison methods.

Semi-supervised Learning Method for Multi-label Tabular Data

GE Zeqing, HUANG Shengjun

Computer Science. 2026, 53 (3): 151-157. doi:10.11896/jsjkx.250600149

Abstract

PDF(2060KB) ( 104 )

References | Related Articles | Metrics

Tabular data is ubiquitous in industrial applications,spanning fields such as medicine,finance,and manufacturing,where each sample is characterized by heterogeneous features.Multi-label classification for tabular data is crucial for capturing the complex,interconnected nature of real-world phenomena,yet obtaining large-scale labeled datasets is often costly.While semi-supervised learning has shown success in image and text data by leveraging unlabeled samples,its application to tabular data remains challenging due to the lack of inherent spatial or semantic structures,making conventional augmentation and consistency-based methods less effective.To address these challenges,this paper proposes a novel semi-supervised learning frameworktai-lored for multi-label tabular data.This approach introduces a structure-preserving data augmentation method that adds Gaussian noise to the feature representation space preserving the original data structure,and a consistency-based regularization technique between samples and their perturbed versions to enhance generalization.Additionally,an attention-based mechanism is developed to selectively aggregate neighborhood information from labeled data,allowing the model to leverage local feature correlations effectively.For unlabeled data,a state-of-the-art pseudo-labeling strategy is employed to enable iterative refinement of model predictions.Extensive experiments are conducted on ten public multi-label tabular datasets,covering various domains to validate the robustness of the proposed method.Results demonstrate the effectiveness of the proposed method,advancing the state of semi-supervised multi-label learning for tabular data.

Prompt-conditioned Representation Learning with Diffusion Models for Semi-supervised Clustering

WANG Yiming, JIAO Min, ZHAO Suyun, CHEN Hong, LI Cuiping

Computer Science. 2026, 53 (3): 158-165. doi:10.11896/jsjkx.250600063

Abstract

PDF(2371KB) ( 107 )

References | Related Articles | Metrics

Current clustering methods enhance performance by jointly learning cluster-friendly representation spaces and cluster assignments.However,they remain fundamentally constrained by static embedding spaces primarily derived from pre-trained visual encoders,where cluster assignments rely on rigid metric systems(e.g.,Euclidean distance or cosine similarity) within the fixed feature space.Inspired by the stable training dynamics and conditional control capabilities of diffusion models,this paper proposes a novel semi-supervised clustering framework.Methodologically,it encodes cluster centers as learnable conditional embedding vectors and constructs a noise-prediction-error-driven generative metric function,transcending the traditional Euclidean linear separability constraints.A two-stage dynamic optimization strategy is designed,integrating supervised pre-training with semantic anchoring and unsupervised adaptation with matching losses to balance intra-cluster compactness and inter-class separabi-lity.Theoretically,based on Rademacher complexity and bounded noise-prediction assumptions,it derives an expected risk upper bound of $\mathcal{O}$(k／n) proving the asymptotic consistency of the proposed method on large-scale data and guaranteeing its generalization capability.Furthermore,it demonstrates that supervised information,through strong convexity constraints and Lipschitz continuity of the denoising network,accelerates the decay rate of the dominant error term to $\mathcal{O}$(1＼/nm_c) elucidating the compression effect of labeled data on hypothesis space complexity.Experimentally,the proposed framework achieves competitive results on benchmark datasets such as ImageNet-10,supported by ablation studies validating the efficacy of key components.

Research on Maximal Directed k-plex Enumeration Problem

HOU Jingle, LI Zhengjun, DAI Qiangqiang, LI Ronghua, WANG Guoren

Computer Science. 2026, 53 (3): 166-172. doi:10.11896/jsjkx.250400086

Abstract

PDF(1795KB) ( 105 )

References | Related Articles | Metrics

The directed edge of a directed graph can represent the direction of a relationship or the transfer of data.It is of great help to introduce and expand some classical dense subgraph models of undirected graphs in dense subgraph mining.Therefore,combining the characteristics of digraphs with the definition of k-plex,the subgraph structure in which the nonoutgoing neighbors and nonincoming neighbors of any vertex in a digraph do not exceed k is called a directed k-plex.Output sensitive algorithms for enumerating maximal k-plex in undirected graphs have been proposed,but they cannot be applied directly to directed graphs.To solve this problem,a recursive enumeration algorithm based on graph decomposition is proposed for maximal directed k-plex enumeration problem.In order to further optimize the efficiency of the algorithm,a pruning strategy based on support points is introduced,and an optimization algorithm based on the upper bound of directed k-plex is provided to terminate some invalid search branches.Experimental results on real graph data show that both the graph decomposition algorithm and pruning optimization have achieved good results.The proposed algorithm has strong practicability in processing real graph data,and can complete the processing of hundreds of groups of real world digraphs in KONECT data set within 2h.

Partial Domain Adaptation Based on Machine Unlearning

WU Jiahao, PENG Li, YANG Jielong

Computer Science. 2026, 53 (3): 173-180. doi:10.11896/jsjkx.250200111

Abstract

PDF(2467KB) ( 98 )

References | Related Articles | Metrics

Domain adaptation aims to transfer knowledge from a label-rich source domain to a label-scarce target domain,thereby improving model performance on the target domain while reducing the need for target domain data annotation.As a more realistic extension,partial domain adaptation(PDA) relaxes the assumption of complete label space sharing between the source and target domains,and handles the case where the target label space is a subset of the source label space.The proposed machine unlearning method helps to address the challenging problem of partial domain adaptation by forgetting outlier-weighted categories.Specifically,the method firstly uses a traditional PDA method as the initialization model,while the category weight mechanism identifies the outlier-weighted categories.Then,it screens the source domain dataset based on the outlier-weighted categories and generates a noisy sample dataset,and then performs unlearning on the model to solve the label space mismatch problem between the source and target domains.Finally,it leverages pseudo-labeling techniques to further align the feature distribution of the target domain,thereby promoting positive transfer.Extensive experiments on the publicly available Office-31 and Office-Home benchmark datasets show that the proposed machine unlearning method not only performs on par with the latest PDA methods,but also significantly outperforms traditional PDA methods.

Type-steered Edge Matching for Heterogeneous Graph Similarity Learning

SANG Shilong, CHEN Kejia

Computer Science. 2026, 53 (3): 181-187. doi:10.11896/jsjkx.250300002

Abstract

PDF(2857KB) ( 111 )

References | Related Articles | Metrics

Graph similarity learning aims to measure the similarity between graphs by learning their structures.Graph similarity learning methods based on graph neural networks are still limited to the node-graph level matching paradigms,failing to perceive the edge-level representation and its contribution to graph structure matching.Moreover,edges in real-world graphs usually have different types,representing different semantic relationships between nodes,which are remain underutilized in cross-graph interaction methods.To address this problem,a type-steered edge matching for heterogeneous graph similarity learning(TEM-HGSL) framework is proposed.Firstly,a heterogeneous graph isomorphism network based on the line graph is designed to better learn edge embeddings.Then,a type-aligned edge matching mechanism is introduced to make better use of the semantic information of edges.Finally,the graph similarity calculation at both the edge and graph levels is realized.Experiments results on four heterogeneous graph datasets show that TEM-HGSL can reduce the mean square error in average of 25.65% compared with the best baseline,effectively achieving fine-grained similarity calculation.

Dual-channel Graph Neural Network Based on KAN

WANG Jinghong, LI Pengchao, WANG Xizhao, ZHANG Zili

Computer Science. 2026, 53 (3): 188-196. doi:10.11896/jsjkx.250600067

Abstract

PDF(4107KB) ( 122 )

References | Related Articles | Metrics

GNNs are specialized models designed for graph data and have been successfully applied to various graph learning tasks such as node classification and link prediction.However,most existing GNN models are based on the message-passing paradigm,which fails to fully capture the multi-dimensional relationships between structural information and feature information of nodes.Additionally,traditional activation functions often lead to information loss and lack interpretability in the models.To address these challenges,this paper proposes a novel Kolmogorov-Arnold Network-based Dual-Channel Graph Neural Network(KDCGNN).KDCGNN employs structural convolution and feature convolution in two separate channels to extract structural and feature information from graphs,generating structural and feature encodings for nodes.These encodings are then fused through concatenation and further transformed using the Kolmogorov-Arnold Network to enhance classification performance and model interpretability.Furthermore,a consistency loss function is introduced to encourage distributional alignment between structural and feature encodings,thereby improving the generalization capability of the model.Experiments on three benchmark citation network datasets(Cora,Citeseer,and Pubmed) demonstrate that KDCGNN outperforms existing baseline methods in node classification tasks.KDCGNN provides a novel approach to improving the interpretability and performance of graph neural networks.

Social Learning Based on Multi-expert Collaboration and Information Interaction

LI Linhao, XU Yanan, DONG Yongfeng, WANG Zhen

Computer Science. 2026, 53 (3): 197-206. doi:10.11896/jsjkx.250100068

Abstract

PDF(3184KB) ( 103 )

References | Related Articles | Metrics

In distributed environments,data heterogeneity manifests as discrepancies in data features.Expert model collaboration suffers from knowledge isolation and improper task allocation,leading to uneven training results among experts,preventing full exploitation of each model’s advantages,and thus limiting overall performance.To address these challenges,this paper proposes MECII.The framework integrates the mixture of experts(MoE) architecture with social learning(SL) principles,optimizing the knowledge sharing and complementarity among experts through four key components:multi-expert collaboration,gating network,adaptive information interaction,and gating selection constraints.This approach effectively resolves issues related to data heterogeneity and expert collaboration in distributed learning.By ensuring precise expert selection and task allocation,MECII facilitates information flow between experts,enhancing each expert’s accuracy when processing specific data,and consequently enhancing the overall model performance.Experimental results demonstrate that MECII significantly outperforms traditional federated learning(FL) baseline methods on the CIFAR-10 and CIFAR-100 datasets.Particularly in scenarios with data heterogeneity,MECII achieves improvements in classification accuracy of 6.69 percentage points and 5.13 percentage points,respectively,compared to the state-of-the-art FedL2P method.Moreover,individual expert accuracy is effectively optimized,validating the framework’s significant advantages in promoting expert collaboration and improving individual accuracy.

Overlapping Community Detection with Graph Regularized Fuzzy Autoencoder

ZOU Xiaoyang, JU Hengrong, CAO Jinxin, MA Xingru, HUANG Jiashuang, DING Weiping

Computer Science. 2026, 53 (3): 207-213. doi:10.11896/jsjkx.250100093

Abstract

PDF(3594KB) ( 115 )

References | Related Articles | Metrics

In complex network analysis,mining community structure is an important and challenging research topic.The existing deep learning-based methods have achieved good results in graph-related tasks,but they rarely deal with community detection tasks,especially overlapping community detection,and fail to fully mine and utilize network topology information.In response to these problems,this paper proposes an overlapping community detection with graph regularized fuzzy autoencode(FAE).Firstly,an autoencoder is employed to encode the network topology into a low-dimensional representation.This is followed by applying fuzzy C-means clustering to generate a fuzzy membership matrix,which is then decoded to reconstruct the network topology.Next,a graph regularization term－designed to characterize structural information within the network－is integrated into the aforementioned autoencoder framework.Subsequently,the autoencoder architecture with the graph regularization forms a stacked autoencoder to derive a deep fuzzy membership matrix.Finally,based on fuzzy set theory,the deep fuzzy membership matrix is utilized to partition overlapping communities.Experimental results on 3 groups of artificial networks and 6 real networks show that the performance of the proposed method evaluates by overlapping normalized mutual information(ONMI),Jaccard index(Jaccard) and F1-Score is superior to that of most of the 7 classical methods,demonstrating its potential in dealing with overlapping community detection problems.

Computer Vision Applications in Rail Transit Systems

ZHAO Binbei, ZHU Li, ZHAO Hongli, LI Yutong

Computer Science. 2026, 53 (3): 214-224. doi:10.11896/jsjkx.250400009

Abstract

PDF(3295KB) ( 119 )

References | Related Articles | Metrics

As the backbone of transportation networks,rail transit systems play a pivotal role in modern society due to their high efficiency and operational reliability.With continuous technological advancements,computer vision technologies have emerged as a critical driver for enhancing rail transit systems toward greater efficiency and dependability.This paper comprehensively examines the current research landscape of computer vision applications in rail transit,evaluates their significant contributions to improving transportation efficiency and safety,and analyzes both the challenges encountered in practical implementations and potential improvement strategies.Through systematic analysis of three primary application domains－station security surveillance,track condition monitoring,and rolling stock status assessment－the study elucidates the implementation frameworks of computer vision technologies while identifying current research trajectories.Finally,the paper provides a forward-looking perspective on development trends,predicting how computer vision will further propel automation and intelligentization in rail transit systems.It also anticipates innovative breakthroughs in this field while ensuring data security compliance,ultimately fostering safer and more sustainable urban transportation ecosystems.

Low-bitrate and Real-time Multiview Video Streaming with 3D Gaussian Splatting

WANG Yizong, NING Hongbo, WANG Haofeng, MA Siwei, GAO Wen

Computer Science. 2026, 53 (3): 225-230. doi:10.11896/jsjkx.250700104

Abstract

PDF(2502KB) ( 127 )

References | Related Articles | Metrics

Multiview videos can offer viewers immersive experiences and enable a variety of applications,but they require times of transmission bandwidth compared to traditional videos.Current multiview coding algorithms mainly leverage redundancy between 2D views and do not consider 3D spatial redundancy.This paper presents a multiview video streaming approach that transforms multiview video content into a compact sparse-view representation to reduce redundancy in 3D space.At the receiver side,the remaining views are synthesized through 3D reconstruction based on this representation.Specifically,this paper proposes a compact multiview video representation based on sparse-views,where the remaining views are synthesized using 3D Gaussian reconstruction and splatting,and a view selection method that selects views to optimize visual quality of synthesized views.Experiments show that the proposed method achieves at least a 44.6% bitrate reduction compared with the baseline and supports end-to-end streaming at over 30 FPS.

Multi-feature Enhanced Association Strategy for Multi-object Tracking

CHEN Yunfang, FANG Qian, LYU Zunwei, ZHANG Wei

Computer Science. 2026, 53 (3): 231-239. doi:10.11896/jsjkx.241100094

Abstract

PDF(3244KB) ( 109 )

References | Related Articles | Metrics

In complex scenarios,multi-target tracking faces problems such as dense target occlusion,nonlinear target motion,poor association matching algorithms leading to identity matching errors,and frequent identity switching.This paper takes ByteTrack as the baseline algorithm and improves its association strategy from three aspects:motion model,weak feature data association,and matching algorithm by fully utilizing existing discriminative features.A multi-objective tracking algorithm with multi discri-minative feature enhanced association strategy is proposed.Firstly,in response to the problem that conventional Kalman filtering is difficult to predict the target position of nonlinear motion,the noise covariance of Kalman filtering is dynamically adjusted using prediction similarity and detection confidence to optimize the motion model and improve the accuracy of target position prediction.Secondly,by integrating a secondary association algorithm,weak feature data association is performed between low-confidence detections and tracks that remain unmatched after the initial association,reducing mismatches between them.Finally,for low confidence detection targets,relative depth is used to decompose the detection targets and trajectories,and a cascaded matching algorithm is used for association,effectively reducing IoU matching collisions and improving the tracking performance of the algorithm in dense occlusion scenes.On the MOT17 and MOT20 test sets,HOTA is 64.5% and 63.2%,respectively,and all evaluation metrics show significant improvement compared to the baseline algorithm.

Smooth Layout Method of Wiring Harness in Collaborative Convergence Point

GAO Jingyi, LUO Ruiming, LUO Yuetong

Computer Science. 2026, 53 (3): 240-245. doi:10.11896/jsjkx.250300101

Abstract

PDF(2722KB) ( 106 )

References | Related Articles | Metrics

With the continuous advancement of electrification and intelligence,the wiring harnesses in products such as aircraft and automobiles have become increasingly numerous and complex,making their layout and installation more challenging.Nume-rous wiring harnesses are installed through a series of convergence points known as clamps,where a single harness may pass through multiple convergence points,and a single convergence point may contain multiple harnesses.If the layout of the wiring harness within the convergence point is not reasonable,it may lead to problems such as crossing and twisting between the wiring harnesses,which does not meet the requirements for smoothness of the wiring harness.Currently,engineers mainly manually adjust the layout of harnesses within convergence points based on experience.However,due to the large number of convergence points and their mutual influence,manual adjustment is not only labor-intensive but also difficult to achieve optimal results.To address this issue,this paper proposes a smoothness-oriented collaborative wiring harness layout algorithm in convergence points.The method firstly transforms the wiring harness layout within multiple convergence points into a set of optimization problems,obtaining a preliminary layout that meets the smoothness requirements through optimization solving.It mathematically models the wiring harness smoothness problem and uses optimization algorithms to obtain the layout of the harnesses within all convergence points,ensuring the smoothness of the harnesses.Then,a compact algorithm for ensuring the layout of the wiring harness is proposed,which maintains the smoothness of the wiring harness and compactly arranges the wiring harness within the convergence point to save installation space and increase the stability of the wiring harness.The proposed algorithm has been applied to the products of our partners,and through the testing of synthetic and actual cases,the results demonstrate the effectiveness of the method.

Student Behavior Detection Method Based on Improved YOLO Algorithm

WANG Xinyu, GAO Donghuai, NING Yuwen, XU Hao, QI Haonan

Computer Science. 2026, 53 (3): 246-256. doi:10.11896/jsjkx.241100165

Abstract

PDF(5717KB) ( 127 )

References | Related Articles | Metrics

In order to solve the problems of large scale variations,serious occlusions,and large computational burden that makes it difficult to popularize on a wide scale for student behavior detection in classroom scenarios,this paper proposes a lightweight student classroom behavior detection method BDEO-YOLO based on the improved YOLOv8.Firstly,dynamic convolution is introduced on the basis of YOLOv8n C2f module in YOLOv8,which enhances the model’s adaptability to complex classroom scenarios and feature expression ability.Secondly,the multi-scale feature fusion ability of the model is optimized by combining Bi-FPN and GLSA,and ELA mechanism is introduced in the Backbone part of the model,which enhances the model’s ability to detect small targets and detailed features.Finally,a lightweight detection head one13 structure is designed to simplify the feature extraction process and significantly reduce the computational burden of the model.Experimental results on the public dataset STBD-08 show that the mAP of the BDEO-YOLO model reaches 92.2%,which is 1.3 percentage points higher than that of the ori-ginal YOLOv8n,and the computational burden is reduced from 8.1 GFLOPs to 4.8 GFLOPs,which is 40.7% lower than the ori-ginal model,and the model size is only 5.7 MB,which verifies the effectiveness of the lightweight design.Validation on the public datasets SCB-Dataset3 and VOC2007 shows that the improved algorithm improves in all performance metrics,verifies the genera-lization ability of the model,and exhibits high robustness in dealing with occlusion,scale change,and illumination change in the classroom.

Efficient Multi-view Stereo Reconstruction Based on Multi-granularity Feature Aggregation and Binary Search

XU Lijun, ZHAO Yujie, ZHAO Min, MA Weixuan, CHEN Kansong

Computer Science. 2026, 53 (3): 257-265. doi:10.11896/jsjkx.250200094

Abstract

PDF(3817KB) ( 121 )

References | Related Articles | Metrics

In deep learning-based multi-view stereo(MVS) reconstruction,cost volume construction faces challenges of high computational complexity and memory consumption.Existing studies often employ cascade architectures or iterative optimization methods to reduce memory usage.However,the coarse-to-fine sampling strategy in cascade structures may lead to the loss of fine-grained details,weakening the perception of critical features.To address this,this paper proposes a novel multi-view stereo network framework based on a cascade structure with binary search and multi-granularity feature aggregation.The proposed framework reduces memory overhead through a cascade architecture while employing a binary search strategy to partition the depth range into multiple candidate regions.A discrete classification method is introduced to compress the depth search space,improving depth retrieval efficiency and lowering memory requirements.Furthermore,this paper proposes a multi-granularity feature aggregation strategy that embeds coarse-grained global semantic information into fine-grained cost volume construction while preserving attention to fine-grained local texture details.By fusing multi-level feature representations and incorporating intra-view adaptive aggregation and view-wise adaptive weighting strategies in the aggregation module,the proposed model enhances the perception of both global structures and local detailed features.Experimental results on the DTU and Tanks & Temples benchmark datasets demonstrate that the proposed method achieves superior point cloud reconstruction quality while maintaining low memory consumption.

Integrate ByteTrack’s EAP-YOLOv8 UAV Marker Point Detection and Tracking

TANG Xinliang, PAN Xiaorun, WANG Jianchao, SU He

Computer Science. 2026, 53 (3): 266-276. doi:10.11896/jsjkx.241100115

Abstract

PDF(4767KB) ( 148 )

References | Related Articles | Metrics

With the development of science and technology,drones are more and more widely used,and the realization of accurate motion capture of drones has become its core technology.When the optical motion capture system detects and tracks the UAV,due to the interference of complex environment,flight speed and other aspects,the Marker point pasted by the UAV will be inaccurate.In order to solve this problem,an improved object detection algorithm EAP-YOLOv8 based on YOLOv8 is proposed to improve the accuracy of Marker point recognition detection.Firstly,a new channel attention mechanism MAP-ECA is constructed in the backbone part,which enhances the global perspective information and the characteristics of different scales,and improves the detection ability of small targets.Secondly,on the basis of the original detection head,the multi-level adaptive feature fusion is used to form a new detection head,D-SASFF,and the multi-scale fusion is used to strengthen the feature information of small targets.Finally,the loss function PIoUv3 is designed,which accelerates the convergence speed of the model and improves the detection ability of small targets.In order to verify the effectiveness of the EAP-YOLOv8 algorithm,experiments are carried out on the self-made dataset,and the results show that the EAP-YOLOv8 algorithm reaches 96.5% and 50.2% on mAP@0.5 and mAP@0.5:0.95,respectively,which is significantly improved compared with other algorithms.On this basis,the tracking accuracy of Marker points is significantly improved by combining the multi-target tracking algorithm ByteTrack,and the tracking experiments are carried out on the public dataset MOT16,and the results show that the new model reaches 37.60%,25.64% and 80.76% on HOTA,MOTA and MOTP,respectively,which is significantly improved compared with the current algorithms,providing an effective way for the subsequent accurate tracking of UAVs.

CA-SFTNet:Skin Lesion Segmentation Model Based on Spatial Feature Transformation and Concentrated Attention Mechanism

ZHANG Wei, LIANG Dunying, ZHOU Wanting, CHENG Xiang

Computer Science. 2026, 53 (3): 277-286. doi:10.11896/jsjkx.250200049

Abstract

PDF(4007KB) ( 121 )

References | Related Articles | Metrics

To address issues such as blurry skin lesion boundaries,noise caused by hair,incomplete segmentation of lesion regions,and significant differences in lesion feature distribution,this paper proposes CA-SFTNet,a U-Net-based algorithm integrating a condensed attention neural block and residual spatial feature transformation.Firstly,feature segmentation during downsampling preserves shallow semantic lesion information.Secondly,condensed attention neural block in skip connections enhances focus on lesion regions by adaptively weighting critical features.Finally,a residual spatial feature transformation module is integra-ted at the network’s tail,enabling adaptive adjustment for spatially heterogeneous regions and enhancing recognition of lesions with heterogeneous feature distributions.Experiments conducted on the ISIC2017 and ISIC2018 datasets demonstrate that CA-SFTNet outperforms the conventional U-Net in skin lesion segmentation.Specifically,it achieves Dice coefficients of 93.12% and 92.36%,representing improvements of 7.15 and 4.81 percentage points over U-Net,respectively.The corresponding IoU values are 82.59% and 82.31%,which constitute gains of 6.23 and 4.45 percentage points.Moreover,when compared with state-of-the-art Transformer-based architectures such as TransUNet and Swin-UNet,CA-SFTNet consistently improves the Dice coefficient by 2～6 percentage points and the IoU by 1.8～4.0 percentage points.These results collectively demonstrate the superiority of the proposed method in skin lesion segmentation and its effectiveness in enhancing segmentation accuracy.

Pedestrian Re-identification Methods Based on Limited Target Data and Deep Learning

FU Yukai, LI Qingzhen, DONG Zhixue, SHI Dongli, ZHAO Peng

Computer Science. 2026, 53 (3): 287-294. doi:10.11896/jsjkx.260100073

Abstract

PDF(3158KB) ( 121 )

References | Related Articles | Metrics

Person re-identification(ReID) has significant application value in cross-camera retrieval scenarios,but deep models often face a significant domain shift problem in real-world deployments.This means that a model trained well on the source domain dataset experiences a sharp performance drop when transferred to a new target camera network.Existing unsupervised domain adaptation methods typically rely on large amounts of unlabeled target domain data for offline clustering and self-training.However,this prerequisite is often difficult to meet in situations involving temporary deployments,privacy constraints,or difficulty in collecting target data in advance.To address this issue,this paper proposes a deep person re-identification adaptation framework based on a small amount of target data.Starting with a pre-trained model in the source domain,it freezes the backbone parameters and introduces only a lightweight,efficient adaptation module for target domain calibration.Simultaneously,it employs a prototype-based stable few-sample decision-making approach,aggregating a small number of labeled target samples into class centers to reduce few-sample noise.Furthermore,it combines prototype classification loss,ranking constraints,and distillation regularization for optimization,balancing target domain adaptability and feature stability.In cross-dataset migration experiments on Market-1501 and DukeMTMC-reID,the proposed method achieves significant improvements in both migration directions:79.68% mAP and 93.10% Rank-1 on Market→Duke,and 76.07% mAP and 93.79% Rank-1 on Duke→Market,with a continuous improvement trend in incremental adaptation rounds.This method can achieve effective and iterative cross-domain adaptation without relying on large-scale target data.

Survey of Table Question Answering Research

WU Xianjie, LI Tongliang, LI Zhoujun

Computer Science. 2026, 53 (3): 295-306. doi:10.11896/jsjkx.250900006

Abstract

PDF(2866KB) ( 116 )

References | Related Articles | Metrics

Tables serve as essential data carriers,capable of efficiently storing large volumes of high-value information.They are widely used across domains such as economics,finance,scientific research,and more.Table question answering(TableQA) aims to automatically derive answers from tabular data in response to natural language queries,representing a key research direction at the intersection of natural language processing and data analysis.Compared to traditional text-based or knowledge-base question answering,TableQA presents greater challenges,as it requires not only natural language understanding but also the interpretation of two-dimensional table structures,numerical computations,and complex logical reasoning.In recent years,the continuous deve-lopment of diverse datasets has driven steady progress in TableQA research.The field has evolved from early rule-based and template-based approaches to statistical learning and neural network models,and more recently,to the integration of pre-trained language models,resulting in consistent performance improvements.Notably,the emergence of large language models(LLMs) has ushered in a new phase of development.Leveraging their strong cross-task generalization and reasoning capabilities,LLMs have accelerated innovation and fostered new research paradigms in TableQA.This paper systematically reviews the evolution and representative methods of TableQA,with a particular emphasis on recent advances enabled by LLMs.It also outlines the key challenges currently facing the field and provides a forward-looking perspective on future research directions.

Review of Speech Disorder Assessment Methods Driven by Large Language Models

XU Cheng, LIU Yuxuan, WANG Xin, ZHANG Cheng, YAO Dengfeng, YUAN Jiazheng

Computer Science. 2026, 53 (3): 307-320. doi:10.11896/jsjkx.250300125

Abstract

PDF(3221KB) ( 129 )

References | Related Articles | Metrics

The growing recognition of speech disorders’ detrimental effects on cognitive development and social adaptability has positioned intelligent assessment systems as a pivotal research priority in language rehabilitation.Conventional assessment approaches,reliant on manual observation and surface-level feature analysis,suffer from inherent limitations including subjectivity,inefficiency,and poor generalizability across diverse scenarios.In contrast,large language models(LLMs)-driven assessment technologies have significantly improved the objectivity and precision of pathological speech detection by integrating multimodal data with deep semantic modeling.This study comprehensively maps the technological evolution in speech disorder evaluation,tracing its progression from acoustic feature extraction to multimodal fusion architectures,with a focused analysis of Transformer-based multimodal integration methods and their groundbreaking applications in cross-linguistic adaptation and real-time intervention strategies.Comparative evaluations of mainstream datasets and metrics highlight LLMs’ superior performance in tasks like speech intelligibility quantification and semantic consistency verification.However,current methodologies encounter persistent challenges,such as inadequate dynamic adaptation of evaluation criteria and unaddressed biases in generative processes.Future research must prioritize the development of dynamically scalable assessment frameworks,leveraging ethical governance mechanisms and cross-modal contrastive learning to overcome high-dimensional semantic consistency modeling barriers.Simultaneously,enhancing clinical validation and privacy-preserving protocols will drive intelligent assessment technologies toward greater precision and fairness.Collectively,these advancements offer methodological blueprints for building adaptable,clinically viable systems,accelerating their scalable deployment in educational support and telemedicine ecosystems.

Few-shot Continuous Toxicity Detection Based on Large Language Model Augmentation

LI Wenli, FENG Xiaonian, QIAN Tieyun

Computer Science. 2026, 53 (3): 321-330. doi:10.11896/jsjkx.250600010

Abstract

PDF(2950KB) ( 104 )

References | Related Articles | Metrics

Toxic speech detection is a challenging problem plaguing online social media.While existing methods can effectively identify common toxic information or toxic information generated through specific perturbation patterns,they face two major challenges:1)Due to the diversity of toxicity types and linguistic expressions,training data cannot cover all samples,leading to a shortage of toxic text data for detection techniques;2)Malicious users in real-world scenarios tend to create new perturbation patterns to deceive text toxicity detectors.How to transfer the model’s detection capabilities for old perturbation patterns to new ones has become an urgent issue to address.To address these issues,this paper proposes a few-shot continuous toxicity detection model based on large language model augmentation.The core idea is to use large language models to augment examples in the training set,then combine continuous learning with toxicity detection techniques to ensure the toxicity detection model can continuously and efficiently detect toxicity in text.Additionally,the model not only achieves more precise understanding of features related to different disturbance patterns but also enhances its adaptability and robustness in the few-shot continuous toxicity detection task.The model is tested on the latest DynEscape dataset,and the results demonstrate that it outperforms existing baseline models,achieving optimal performance.

Embedding Model of Knowledge Graph via Jointly Modeling Ontology and Instances

QIN Jing, LI Guanfeng, CHEN Yuyin, XIAO Yuhang

Computer Science. 2026, 53 (3): 331-340. doi:10.11896/jsjkx.250200101

Abstract

PDF(1698KB) ( 92 )

References | Related Articles | Metrics

Knowledge graph embedding provides a more powerful knowledge representation input to machine learning models by projecting entities and relationships into a continuous low-dimensional vector space,thereby supporting more knowledge graph application scenarios.In recent years,researchers have tried to use the potential semantic information between ontology and instance in knowledge graph to enhance the embedding of knowledge graph.However,they fail to effectively integrate the hierarchical structure of concepts and the specific information of instances,and ignore the transitivity between isA relationships,resulting in limited performance and generalization ability of the models when dealing with long-tail entities in the knowledge graph.In order to solve the above shortcomings,this paper proposes a knowledge graph embedding model(Representation Learning of Knowledge Graph via Jointly Modeling Ontology and Instances,JMOI),which integrates ontology and instance.By introducing self-attention mechanism,this model captures the complex semantic relationship between concepts and instances,and adds a learnable parameter to adjust the neighborhood range of concept embedding,so as to distinguish the hierarchical information of diffe-rent concepts.The transitivity of isA relationship is modeled.Experimental results on the YAGO26K-906 and DB111K-174 datasets show that JMOI achieves the best performance in most cases compared with the prior art,with a maximum improvement of 6.5% in the link prediction Hits@1 and 6.9% in the Recall in triple classification compared with the suboptimal model.

Enhanced Multi-turn Machine Reading Comprehension for Aspect Sentiment Triplet Extraction

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

Computer Science. 2026, 53 (3): 341-350. doi:10.11896/jsjkx.250300039

Abstract

PDF(3003KB) ( 103 )

References | Related Articles | Metrics

ASTE aims to simultaneously extract aspects,their corresponding opinions,and sentiment polarities from text.It is an emerging and challenging task in aspect-level sentiment analysis.Among existing methods,those based on multi-turn machine reading comprehension have effectively achieved sentiment triplet extraction,but they still exhibit certain limitations.Firstly,the single text feature in multi-turn reading comprehension struggles to adapt to specific subtasks.Secondly,the global self-attention mechanism lacks focus on syntactically more important words and assigns higher attention weights to less significant words.To address these issues,this paper proposes an enhanced multi-turn machine reading comprehension(EMT-MRC) method,which designs a bidirectional attention flow in each turn of reading comprehension to construct the interaction between text and questions,thereby obtaining task-specific text representations.Additionally,dependency syntactic relations are integrated into the Transformer encoder,which constrains the model’s attention distribution through dependency distances,thereby enhancing the model’sfocus on the grammatical aspects of sentences.Experiments on two groups of datasets demonstrate the effectiveness of the proposed method.

Large-scale Multi-objective Evolutionary Algorithm Based on Objective Similarity and Dual-EndVariable Guided Search

YANG Changhao, QIN Jin, WANG Hao

Computer Science. 2026, 53 (3): 351-365. doi:10.11896/jsjkx.250200091

Abstract

PDF(3680KB) ( 110 )

References | Related Articles | Metrics

Large-scale multi-objective optimization problems (LSMOPs) involve a large number of decision variables,resulting in expansive search spaces that make them challenging for traditional evolutionary algorithms to find good solutions efficiently within limited resources.To address this,a large-scale multi-objective evolutionary algorithm based on objective similarity and dual-end variable guided search (LMOEA/OS-DES) is proposed.LMOEA/OS-DES includes three strategies.The first strategy is the co-evolution of multiple swarms driven by objective similarity in order to quickly obtain solutions that reflects the distribution characteristics of Pareto optimal solution set.The second strategy is to design various variable grouping schemes based on the distribution characteristics of elite solutions in the decision space,so as to adapt to the differences between the distribution of optimal solutions for different objective vector directions.Combined with the grouping schemes,the dual-end variable guided search generates new solutions with distribution characteristics similar to the previous elite solutions,which enables it to adopt larger variations than the previous strategy,explore more regions faster,and accelerate the optimization of convergence and diversity.In the final strategy,it uses the competitive swarm optimization to explore the regions around elite solutions,so as to rapidly optimize diversity.Comparative experiments with eight other competitive algorithms on LSMOP and UF with dimensions ranging from 100 to 5 000 demonstrate that LMOEA/OS-DES has strong advantages.

Optimizing Probabilistic Choice for Solving SAT Problems

JIA Shuheng, FU Huimin

Computer Science. 2026, 53 (3): 366-374. doi:10.11896/jsjkx.241200211

Abstract

PDF(1989KB) ( 99 )

References | Related Articles | Metrics

In stochastic local search algorithms for SAT problems,mainstream variable decision strategies rely on probabilistic variable selection,such as the probSAT solver,which determines selection probabilities by calculating the break value of vari-ables.However,this approach is prone to falling into local optima,particularly underperforming in application-oriented problems.To address this,a variable decision method incorporating a configuration detection strategy is proposed,dynamically adjusting the variable selection probability function.When the environment remains unchanged,variables with lower break values are prioritized,enhancing global optimization capabilities.To tackle the high scanning overhead of long clauses,an important neighbor array strategy is introduced,incorporating highly active variables into the array to reduce computational complexity.Additionally,a restart mechanism is designed to leverage the advantage of probSAT in rapidly reducing the number of unsatisfied clauses during the initial phase,avoiding the global repeated flipping phenomenon in later stages,thereby improving solving efficiency.The improved probSAT_PCCR solver demonstrates significant performance enhancements in long-unsolved mathematical application problem tests,solving 142 more cases than the original probSAT,with a performance improvement of 546.1%.In practical application problem tests conducted by the FCC,it solves 1 596 more cases,with a performance improvement of 33.5%.In summary,the enhanced probSAT solver,through multiple strategic improvements,achieves substantial performance gains in solving application-oriented SAT problems,demonstrating significant application value.

Prediction Method of RNA Secondary Structure Based on Transformer Architecture

YU Ding, LI Zhangwei

Computer Science. 2026, 53 (3): 375-382. doi:10.11896/jsjkx.250100005

Abstract

PDF(3822KB) ( 110 )

References | Related Articles | Metrics

RNA secondary structure prediction is a core problem in bioinformatics,and recent advancements in deep learning have significantly propelled progress in this field.However,existing methods still face limitations in prediction accuracy and reliance on external prior models,which may compromise the robustness and generalization capabilities of these models.To address these issues,this paper proposes a Transformer-based model for RNA secondary structure prediction.The model designs dual feature encoding pathways,generating sequence features through linear embedding and one-hot encoding,and efficiently fuses these two feature representations using a cross-attention mechanism.During the feature extraction phase,the model employs an improved architecture combining Swin-Transformer and U-net(Swin-Unet) to achieve deep-level feature extraction,ultimately producing a pairing probability matrix for RNA secondary structures.Experimental results show that the proposed model achieves over 3% higher F1-scores than other models on multiple benchmark datasets without relying on prior information from external models.This study provides a novel solution for RNA structure prediction and highlights the promising potential of Transformer architectures in biological sequence analysis.

Multi-task Learning-based Ophthalmic Video Feature Fusion and Multi-dimensional Profiling

DU Jiantong, GUAN Zeli, XUE Zhe

Computer Science. 2026, 53 (3): 383-391. doi:10.11896/jsjkx.260200058

Abstract

PDF(2912KB) ( 113 )

References | Related Articles | Metrics

To address challenges in profiling ophthalmic videos on social networks,such as the low discriminability of visual features,the colloquial nature of text descriptions,and multimodal semantic heterogeneity,this paper proposes an OVP(Ophthalmic Video Profiling) method based on multi-task learning.The proposed method aims to mine multi-dimensional features with medical semantic value from unstructured video and text streams to facilitate precise video representation.In the OVP framework,a pre-trained deep residual network is employed to extract high-dimensional visual representations from keyframes,capturing fine-grained features specific to ophthalmic imagery.To overcome the sparsity of professional semantics in social media text,a method for extracting textual features from ophthalmic videos based on an ophthalmic knowledge graph is proposed,which retrieves and fuses external entity annotations and related knowledge before encoding via BERT.Subsequently,a cross-modal attention fusion mechanism is designed to dynamically calculate interaction weights between visual and textual features,achieving deep alignment between visual information and medical semantics.Furthermore,a multi-task joint optimization and ophthalmic multidimensional profiling is constructed to jointly train three sub-tasks:disease classification,popularity prediction,and content quality assessment,utilizing shared information to enhance model generalization.Experiments conducted on a real ophthalmic video dataset demonstrate that the OVP method significantly outperforms existing baseline methods in terms of disease classification accuracy,heatmap prediction,and quality assessment performance for ophthalmic videos.The experimental results validate the effectiveness of the OVP method in feature fusion and multidimensional profiling of complex ophthalmic videos.

A Serverless-based Approach to Fast Measurement of Network Bandwidth

YANG Ruoxuan, JIN Feiyu, QU Lianwei, ZHOU Zijie, ZHENG Qibin, LI Zhenhua

Computer Science. 2026, 53 (3): 392-399. doi:10.11896/jsjkx.250600011

Abstract

PDF(2098KB) ( 108 )

References | Related Articles | Metrics

Next-generation mobile networks,represented by 5G/6G and WiFi 6/7,have substantially increased access bandwidth(i.e.,network speed).At the same time,they have also amplified fluctuations in network performance-such as throughput,latency,and packet loss-thereby prolonging speed tests,increasing traffic overhead,and degrading user experience.More importantly,this volatility makes it difficult to meet the pressing requirement for low-cost and highly real-time networking in microkernel-based ubiquitous computing systems(e.g.,MINIX,QNX,and seL4),where task management and inter-process communication critically depend on timely and lightweight network support.Although serverless computing built on cloud functions and cloud containers offers a potential solution path,existing rapid bandwidth measurement methods(e.g.,Swiftest) remain heavily constrained by the long-tail traffic effect,leading to considerable traffic waste.To address this challenge,this paper propose a serverless-based approach to fast measurement of network bandwidth.We first analyze the causes of the long-tail effect in prior approaches by combining transport-layer mechanism analysis with real-world measurement data.Building on these insights,we design a bursty fast-start transmission mechanism that decomposes the conventional smoothly ramped packet-sending strategy into multiple rounds of short-duration burst transmissions.With dynamic feedback from the client,the sender regulates its transmission rate in real time to shorten measurement duration and improve estimation accuracy,thereby avoiding long-tail traffic waste induced by delayed control feedback.Experiments across multiple representative network scenarios show that,compared with Swiftest,the proposed method reduces server-side transmitted traffic by 85% and shortens the average measurement time to 1.6 se-conds.These gains significantly alleviate server resource pressure and reduce client data consumption,while exhibiting strong engineering deployability in ubiquitous computing environments.

Connectivity and Diagnosability of Data Center Network SWCube

ZHANG Xinfan, CHENG Baolei, FAN Jianxi, WANG Yan

Computer Science. 2026, 53 (3): 400-410. doi:10.11896/jsjkx.250400096

Abstract

PDF(2258KB) ( 94 )

References | Related Articles | Metrics

The growing size of servers necessitates greater efficiency in data center networks.To address this,researchers have introduced various new data center networks to enhance traditional configurations,which often suffer from low reliability,high deployment costs,limited bandwidth,and numerous other drawbacks.SWCube is an innovative data center network designed based on the generalized hypercube,achieved by replacing the nodes of the generalized hypercube with switches and incorporating one dual-port server into each edge connecting two switches.Its logical graph corresponds to the line graph of the generalized hypercube.SWCube has demonstrated numerous advantages,including high bisection width,ability to accommodate many servers,scalability,etc.However,its reliability remains unexamined.The reliability of a network is divided into fault tolerance and fault diagnosis capability,typically assessed by the connectivity and diagnosability obtained from system-level diagnosis under various models.To this end,a recursive definition of the SWCube logical graph is presented.According to this definition,the connectivity of the SWCube logical graph is given as $\sum_{i=1}^{r}\left(m_{i}-1\right)$ for the case where r=2 and m₂=2,and as $2 \sum_{i=1}^{r}\left(m_{i}-1\right)-2$ where r≠2 or m₂>2.Based on this,the diagnosability of SWCube under the PMC model and the MM^* model is derived.

Research Review of Application-based Covert Channel

CHANG Huiyan, HU Hongchao, ZHOU Dacheng, XU Depeng, CHENG Guozhen

Computer Science. 2026, 53 (3): 411-423. doi:10.11896/jsjkx.250400047

Abstract

PDF(2092KB) ( 99 )

References | Related Articles | Metrics

Recently,tunneling covert information in the network information flow generated by secure and trusted applications has become a prevalent technique for building covert channels.With the development of covert channel technology,a variety of application-based covert channel systems have been proposed.However,existing reviews predominantly focus on the principles and definitions of the application-based covert channel technologies,omitting detailed explanations of the classification criteria and the strengths and weaknesses of various systems.Consequently,this paper gives a comprehensive and profound review of the field of application-based covert channels.From the perspective of the underlying applications,these covert channels are classified into two categories:those based on multimedia streaming applications and those based on real-time online games.Further,based on the different stages of embedding covert data,the covert channels based on multimedia streaming are subdivided into two types of methods:embedding in the original multimedia stream and embedding in the compressed multimedia stream.In addition,by conducting an in-depth analysis of the unique advantages and potential problems of each method,a multi-dimensional comparative analysis of covert channels based on multimedia flow is established,aiming to reveal the characteristics and differences of various covert channel technologies.Building on the prior research,the prevailing challenges and prospective development trajectories in this field are delineated.

Dual-channel Source Code Vulnerability Detection Model Based on Contrastive Learning

SONG Jianhua, HE Jiawei, ZHANG Yan

Computer Science. 2026, 53 (3): 424-432. doi:10.11896/jsjkx.250200124

Abstract

PDF(2365KB) ( 89 )

References | Related Articles | Metrics

As software vulnerabilities continue to increase,system security is facing severe challenges.Source code vulnerability detection can identify potential security threats in software applications during the development phase,which is crucial for ensuring the security of software applications.Currently,the mainstream method for source code vulnerability detection is based on deep learning models.However,many existing deep learning models rely only on a single form of features and fail to fully explore both the global and local information in the source code semantics.Additionally,these models often overlook the differences and similarities between different samples,leading to poor performance when handling complex vulnerability patterns,with high false positive and false negative rates.To address these issues,a dual-channel source code vulnerability detection model based on con-trastive learning is proposed.This model uses different channels to separately extract global and local features from the source code semantics and introduces contrastive learning to allow the model to learn the similarities and differences between different samples,thereby optimizing the feature extraction process.Experimental results show that this model shows significant improvements in recall and F1 score on the real-world vulnerability datasets,Devign and Reveal,compared to the baseline models.The average improvement is 14.65 percentage points and 6.30 percentage points on Devign,and 31.18 percentage points and 22.44 percentage points on Reveal.

Data Distribution Matching for Monolithic Firmware Base Address Identification

CAI Ruijie, JIA Fan, YIN Xiaokang, ZHAO Fangfang, LIU Shengli

Computer Science. 2026, 53 (3): 433-442. doi:10.11896/jsjkx.250400026

Abstract

PDF(4738KB) ( 127 )

References | Related Articles | Metrics

Base address identification for monolithic firmware serves as the foundation for firmware security research.Current methods and related tools suffer from low identification rate,poor performance,and high resource consumption.To address these limitations,this paper proposes a new data distribution matching-based method for monolithic firmware base address identification.The method firstly calculates the effective character density of each firmware section,then partitions the firmware into text segments and non-text segments based on this density metric.Then,it comprehensively extracts string constant data from the text segments.By identifying and parsing register load instructions,this method retrieves absolute address data embedded in the firmware.These absolute addresses are clustered according to both their destination functions and the source registers used prior to function invocation.The base address is ultimately determined by matching the distribution patterns between the absolute address clusters and string constant data within the firmware,thereby establishing their correspondence and enabling accurate base address resolution.Experimental results demonstrate that the proposed data distribution matching-based firmware base address identification method significantly outperforms existing approaches in recognition efficiency,achieving a 100% success rate on a test set comprising 30 firmware samples.

Transformer-based Domain Adaptation Method for IoT Traffic Intrusion Detection

ZHU Feng, YE Zongguo, LI Peng, XU He

Computer Science. 2026, 53 (3): 443-452. doi:10.11896/jsjkx.241200167

Abstract

PDF(2076KB) ( 99 )

References | Related Articles | Metrics

With the proliferation of IoT devices,intrusion detection systems(IDS) are essential to safeguard IoT networks from malicious attacks.However,the scarcity of IoT-specific data limits the effectiveness of traditional methods,while existing domain adaptation approaches often rely on coarse alignment,overlooking intrinsic semantic properties and lowering feature discriminabi-lity.To address these issues,this paper proposes a semi-supervised domain adaptation model,named TDAIID.This model aligns NI domain and II domain at domain,class,and sample levels.The cross-attention mechanism ensures fine-grained feature alignment by focusing on similarities between same-class samples in the source and target domains.Multiple geometric semantic alignment is semantically aligned from both domain-level and class-level perspectives,facilitating the cross-attention mechanism in learning richer and more accurate knowledge from the source NI domain.To fully exploit unlabeled target data,a dynamic center-aware pseudo-labeling algorithm is proposed to improve pseudo-label accuracy and mitigate negative transfer caused by mislabe-ling.Experiments on several widely-used intrusion detection datasets demonstrate that the TDAIID model outperforms state-of-the-art baseline methods,showcasing its superior performance on IoT intrusion detection.

Deep Learning-based Neural Differential Distinguishers for GIFT-128 and ASCON

SU Ruitao, REN Jiongjiong, CHEN Shaozhen

Computer Science. 2026, 53 (3): 453-458. doi:10.11896/jsjkx.250600176

Abstract

PDF(2363KB) ( 101 )

References | Related Articles | Metrics

As a critical method for block cipher security evaluation,differential cryptanalysis distinguishes ciphers from random permutations by analyzing plaintext difference propagation during encryption.Traditional approaches struggle with complex cryptographic algorithms,while deep learning offers new cryptanalysis perspectives increasingly applied in recent years.To enhance the security evaluation of block ciphers,this paper proposes a neural differential distinguisher construction method that integrates traditional differential analysis with deep learning.For dataset construction,a triplet input format comprising multiple ciphertext pairs is adopted to preserve differential characteristics and capture cross-ciphertext-pair correlations.The network architecture builds upon Convolutional Neural Networks(CNNs) and incorporates residual shrinkage networks to form a deep expansion structure with a multi-scale feature fusion mechanism.Experiments on GIFT-128 and ASCON-PERMUTATION demonstrate significant improvements:For GIFT-128,the highest accuracy of 6-round and 7-round distinguishers reaches 99.70%(an improvement of 9.30%) and 95.47%(an improvement of 13.09%),respectively.For the 4-round analysis of ASCON,the highest accuracy achieves 53.54%.These results validate the effectiveness of the deep learning approach in cryptographic security analysis.

Efficient Data Sharing Scheme with Integrity Auditing Functions in Cloud Storage

ZHANG Yuhang, CHANG Jinyong, YANG Luyao, XU Maozhi

Computer Science. 2026, 53 (3): 459-468. doi:10.11896/jsjkx.241200102

Abstract

PDF(2072KB) ( 104 )

References | Related Articles | Metrics

With the popularity of cloud storage technology,the accompanying security is becoming more and more prominent:cloud servers may lose users’ stored data due to failures or external attacks,while the process of sharing data based on cloud sto-rage may also face the risk of unauthorized access by malicious users.Existing research mostly focuses on the implementation of a single security feature in cloud storage environments.In this paper,secure access control of data sharing process is accomplished on the basis of security audit of cloud storage data integrity.In the process of data integrity auditing,identity-based homomorphic authentication technology is used to generate tags for the stored data,and the user can be informed of the integrity of the stored data by verifying the aggregated tags returned by the cloud server,thus solving the problem of accidental loss of cloud storage data.In the data sharing phase,the hybrid form of attribute-based encryption and symmetric encryption can not only reduce the computation,communication and storage overhead of outsourced data,but also achieve the control of privilege management for unauthorized users,thus solving the problem of balancing high efficiency and privilege management during data sharing.The performance analysis shows that thedesigned system has low computation and communication overheads as well as storage redundancy in both data integrity auditing and sharing processes,which provides new research ideas for secure data storage and efficient sharing in cloud storage environment.