Computer Science

Neural Network Acceleration Architecture Based on RISC-V Instruction Set Extension

CAI Chenghuan, WANG Yipin, XU Jiabin, ZHANG Fengzhe, ZHOU Xuegong, CAO Wei, ZHANG Fan, YU Xinsheng

Computer Science. 2025, 52 (12): 1-8. doi:10.11896/jsjkx.250600014

Abstract

PDF(2143KB) ( 74 )

References | Related Articles | Metrics

To address the current shortcomings of RISC-V-based neural network accelerators in accelerating matrix computations and nonlinear operations within Transformer-based models,a neural network acceleration architecture based on RISC-V instruction set extension,named Taurus,is proposed.This architecture introduces matrix instruction extensions tailored to the characte-ristics of Transformer models and employs a systolic array to perform matrix multiply-accumulate operations.To accelerate nonlinear computations,vector instruction extensions are added,along with the design of specialized vector units to efficiently compute operations such as LayerNorm and Softmax.To ensure balanced data supply,memory access instruction extensions are optimized to provide sufficient data throughput to the matrix and vector computation units.The instruction set extensions adopt a scalar register expansion approach,embedding operand data information directly into the registers.This increases the addressing space and reduces the number of instructions required for large-scale data computations.The Taurus neural network accelerator architecture is cycle-accurately simulated on the Gem5 platform.Compared with the open-source accelerator Gemmini,Taurus achieves an 80% improvement in systolic array utilization during general matrix multiplication.For inference tasks on ResNet50 and BERT models,Taurus delivers 1.3× and 31.3× speedups respectively over Gemmini.Compared with the baseline RISC-V,Taurus achieves 1 467× and 4 513× performance improvements respectively.

AFL-VTest:Fuzzing Framework for Aerospace Embedded Software

WANG Shuai, HUANG Chen, JIANG Yunsong, XIAO Xi, WANG Guanlin, YU Tingting, XU Qizhen

Computer Science. 2025, 52 (12): 9-17. doi:10.11896/jsjkx.250400144

Abstract

PDF(2691KB) ( 35 )

References | Related Articles | Metrics

The reliability of aerospace embedded software is a critical determinant of space mission success.Fuzzing has become the mainstream method for defect detection and vulnerability discovery today,and has achieved significant success in the field of software security.The research on fuzzing methods for aerospace embedded software has profound significance for enhancing the reliability of such software and promoting the progress of aerospace technology.Therefore,this paper proposes AFL-VTest,a fuzz testing framework specifically designed for aerospace embedded software.It integrates a streamlined source code instrumentation method and a novel checksum-fixing algorithm tailored to address limited memory resources and the prevalence of checksum verifications in embedded systems.Evaluation experiments conducted on multiple sample programs and practical aerospace embedded software demonstrate the effectiveness of the proposed instrumentation method and checksum fixing algorithm.Finally,AFL-VTest successfully uncoveres three previously undetected defects within the actual aerospace embedded software projects,thus verifying the effectiveness and practical value of the proposed method in bolstering the safety and reliability of aerospace systems.

Automated Program Repair Based on Perturbing and Freezing Pre-trained Model

ZHANG Lizheng, YANG Qiuhui, DAI Shengxin

Computer Science. 2025, 52 (12): 18-23. doi:10.11896/jsjkx.241100182

Abstract

PDF(1837KB) ( 29 )

References | Related Articles | Metrics

With the increasing complexity of software,the scale and complexity of program defects are also increasing.Program defects not only consume a large amount of development costs but also lead to real-world security issues.Existing program repair methods generally suffer from poor repair effectiveness and high training costs.To address these issues,this paper proposes an automatic program repair method based on perturbation and freezing of pre-trained models.By adding noise to the model parameters through a matrix-based perturbation method,it alleviates the overfitting problem of pre-trained models on the program repair task during fine-tuning.Furthermore,freezing the encoder in the pre-trained model reduces the model’s training time and computational resource consumption.Additionally,the checkpoint ensemble strategy is adopted to enhance the model’s repair effectiveness.Experiments on 40 Python programs in the QuixBugs dataset demonstrate that the proposed method has significant advantages in reducing model training time and computational resource consumption,as well as in repair effectiveness.It only requires training 41.62% of the parameters of the original model,reduces training time by 39.16%,and can repair 70% of the defects in the dataset,demonstrating the diversity of the repaired defect types.

SCADE Model Checking Based on Implicit Predicate Abstraction and Property-directedReachability

ZHANG Cong, CHEN Zhe, WANG Huijie, WEI Yiyang

Computer Science. 2025, 52 (12): 24-31. doi:10.11896/jsjkx.241100062

Abstract

PDF(1609KB) ( 56 )

References | Related Articles | Metrics

SCADE is widely used in critical industries related to life safety,such as aerospace,nuclear power plants,rail transit,and medical equipment.Applying model checking to these safety critical areas can effectively ensure the safety of the system.At present,there is relatively little research on SCADE model checking.Most studies are based on program translation and use other simpler language model checking tools to complete verification.However,few tools that have implemented the entire process of model checking for SCADE programs have low verification efficiency.This paper proposes a model checking algorithm based on implicit predicate abstraction and property-directed reachability(IAPDR),which is parallelly integrated into existing model che-cking tools(PSMC) for SCADE programs.The tool implements the entire process of analysis,modeling,and model checking for SCADE programs.In addition,the correctness of the proposed algorithm is theoretically proven,and the effectiveness and performance of the expanded tool(PSMCWI) are evaluated through experiments.Compared with traditional BMC,K-Induction,and CEGAR algorithms,IAPDR has the highest verification success coverage and the lowest total verification time on the benchmarks.Compared with the native PSMC tool,PSMCWI can verify 139 more SCADE programs on the benchmarks,with a 15.1% increase in successful verification coverage and a 43% reduction in total verification time.The results of the comparative experiment with JKind show that IAPDR can correctly do model checking on SCADE programs.Compared with the method of translating SCADE models into Lustre models and using JKind to do model checking on Lustre models to achieve model checking on SCADE programs,PSMCWI has higher efficiency.

Cross-procedure Feature Envy Detection Supporting Type-sensitive Scenarios

LI Jianhao, BAI Yaoyao, MI Jie, ZHANG Yingzhou, CAO Wenlong, WANG Dong, WANG Gang

Computer Science. 2025, 52 (12): 32-39. doi:10.11896/jsjkx.241200007

Abstract

PDF(1763KB) ( 33 )

References | Related Articles | Metrics

The existence of feature envy phenomena can affect the stability and maintainability of the system.Current feature envy detection methods fail to consider the sensitivity of object types,leading to low detection accuracy.To address this issue,an inter-procedure feature envy detection method based on higher-order function is proposed.The method stores the computation process of the local reference ratios within a procedure with parameters as a summary in the form of higher-order function,according to predefined feature envy measurement rules.During inter-procedure detection,it retrieves the higher-order feature envy detection summary of the target method at the method call site.Based on the correlation between formal and actual parameters,the actual types of the formal parameters are substituted into the summary to compute the final local reference ratios set,which is used to detect the presence of feature envy and corresponding envy sets.The paper integrates several Java projects as a benchmark test set and selects the IntelliJDeodorant and IDE Inspection tools for comparison experiments.Results show that the proposed methodimproves detection accuracy for envy instances by 16.6% over IDE Inspection and by 1.3 times over IntelliJDeodorant.In terms of envy sets detection accuracy,it improves by 37.2% over IDE Inspection and by 1.6 times over IntelliJDeodorant.

Distributed Automated Testing for Android Applications Based on Reinforcement Learning

SONG Rirong, CHEN Qinwen, CHEN Xing

Computer Science. 2025, 52 (12): 40-47. doi:10.11896/jsjkx.241100054

Abstract

PDF(2728KB) ( 20 )

References | Related Articles | Metrics

Android applications have become deeply integrated into various aspects of people’s daily lives.However,ensuring the correctness of these applications remains a highly challenging task.Traditional testing methods primarily rely on manual operations,while automated testing technologies,despite advancements,still require improvements.The continuous iteration of Android applications to enhance performance and meet functional requirements has led to increased application complexity and an explosive growth in state combinations.Testing the core aspects of Android applications involves exploring deep-seated failures under complex user interactions,but the search space for these failures is vast,necessitating substantial time investment for thorough testing.In recent years,researchers have begun employing reinforcement learning to test Android applications by adjusting exploration strategies based on rewards obtained during the interaction between agents and the Android applications.However,existing work has been limited to testing with a single device,significantly constraining testing efficiency.To address these challenges,this paper proposes a distributed automated testing framework for Android applications based on reinforcement learning,named DistributedAndroidExplore(DAE).DAE utilizes multiple agents to concurrently conduct reinforcement learning-based testing on applications and periodically iterates to aggregate the cumulative learning experiences of each agent,thereby enhancing testing efficiency.DAE is evaluated on 10 real-world Android applications.The results indicate that,in most cases,DAE surpasses the compared benchmark algorithms in terms of fault detection rate and code coverage.Furthermore,DAE demonstrates notably higher testing efficiency,with performance improvements ranging from 16.5% to 34.3%.

Transition and Verification Method from RSML^-e to Lustre Model for Flight Mode Transition

WANG Zhiyi, HU Jun, XU Heng

Computer Science. 2025, 52 (12): 48-59. doi:10.11896/jsjkx.250600027

Abstract

PDF(2496KB) ( 31 )

References | Related Articles | Metrics

The automatic flight system is the core system of flight control for modern large aircraft.The flight guidance system,as the core subsystem of the automatic flight system,manages and controls the mode transition of the automatic flight system.The different flight stages of an automatic flight system are essentially flight mode transition,which determine the flight safety of the aircraft.However,flight mode transition has the essential characteristics of a multi-dimensional complex static structure that is coupled and compatible,as well as a multi-level dynamic combination of interactive cooperation and transitional switching.Therefore,ensuring the correctness of flight mode transition is of crucial importance.The model-driven software modeling method models the flight mode transition requirements as semi-formal or formal models,thereby analyzing and verifying the properties satisfied by the model.The existing methods face two challenges:1)The modeling from natural language requirements to semi-formal requirement models is mostly done manually,and there are differences among the requirement models established by different modeling languages;2)The semi-formal requirements model cannot be directly used for model checking.It needs to be transitioned into the input model of the model checking tool,and the verification efficiency of different verification tools also varies.Based on the RSML^-e requirement model of automatic flight mode transition,this paper proposes a systematic method for transitioning the RSML^-e model into the Lustre synchronous data flow language.Firstly,mapping rules are constructed from multiple dimensions such as data types,variables,logical phrases,AND-OR tables,and macros.It transitions the elements in the RSML^-e model one by one into the forms supported by the Lustre synchronous data stream language,and reduces the number of variables when describing the safety properties.Secondly,after the model transition,the Lustre model and safety properties obtained from the transition are input into the Jkind model checking tool for verification.Based on the various optimization techniques built into the Jkind model checking tool,the problem of state space explosion during the model verification process is better alleviated,and efficient verification of large-scale models is achieved.Ultimately,the safety properties related to the mode transition requirements of the automatic flight system are successfully verified through this process,ensuring the operational reliability of the system under various working conditions.

Multiscale Sunspot Number Forecasting Based on Decomposition and Integration

ZHAO Yuxuan, YU Dingfeng, LI Dongxue, XU Yidong, LI Beiming

Computer Science. 2025, 52 (12): 60-70. doi:10.11896/jsjkx.241100011

Abstract

PDF(6677KB) ( 26 )

References | Related Articles | Metrics

Solar activity exerts a direct influence on the heliosphere environment and life on earth.SN represents one of the most crucial and frequently predicted indices of solar activity.Enhancing the accuracy of SN predictions can provide more reliable data support for climate models,which is of great significance for understanding the solar activity cycle.This paper proposes a multi-scale SN sequence prediction model,which combines Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN),a hybrid neural network and an attention mechanism.The method employs three distinct datasets:the Daily Total SN from 1818 to 2024,the Monthly Mean Total SN from 1749 to 2024,and the Yearly Mean SN from 1700 to 2023.Given the non-stationary,non-Gaussian and non-linear nature of the SN time series,the CEEMDAN method is initially employed to decompose the components of solar activity changes on various time scales into a number of sub-series with different frequencies.These sub-series are then combined with the original series as a reinforced feature set,thereby enhancing the model’s ability to characterize the changes in solar activity.TCNs are then employed as the primary means of feature extraction,followed by the incorporation of BiLSTM to capture the long-term dependence of the time series.Additionally,Attention mechanisms are introduced to dynamically identify and weight the key temporal features in the sequence.Ablation experiments are conducted on three datasets concurrently,and the results demonstrate a notable synergy between the modules of the proposed model.A comparison of the existing models on this basis reveals that the prediction accuracy of each dataset has been enhanced.The proposed model is employed to predict SN,resulting in the acquisition of three distinct frequencies of yearly,monthly,and daily prediction outcomes.These are subsequently integrated as multi-timescale features to generate the final prediction results.The results indicate that solar activity has shown a significant increasing trend in 2025 and is expected to reach the peak of Solar Cycle 25 within this year,with an estimated annual mean SN of 233.9.

Cross-regional Battery Swapping Station Deployment Algorithm Based on Intelligent E-scooterFlow Generation

CHEN Jiayi, GU Chengyi, ZHOU Jihua, ZHAO Tao, WANG Shuangchao, ZHU Mingxing, XIANG Chaocan

Computer Science. 2025, 52 (12): 71-80. doi:10.11896/jsjkx.250200010

Abstract

PDF(4376KB) ( 13 )

References | Related Articles | Metrics

Battery swapping services are becoming increasingly popular as intelligent E-scooters emerge as a primary mode of short-distance transportation.As a result,companies providing battery swapping services are expanding their operations in urban regions.During expansion,companies prioritize deploying battery swapping stations in regions with higher user activity,as indicated by intelligent E-scooter flow.However,the lack of operational data in new regions makes data-driven deployment optimization approaches less effective,creating a dilemma of “data scarcity-deployment difficulty-lack of data acquisition”.To address this issue,this paper proposes a cross-regional battery swapping station deployment algorithm based on intelligent E-scooter flow ge-neration.Firstly,a regional flow generation model based on a denoising diffusion probabilistic model is constructed to capture the spatiotemporal characteristics of flow and generate synthetic flow data for target regions using data from already deployed regions.Then,the generated flow is incorporated into a deployment optimization problem,modeled to maximize the overall benefit of station deployment.Finally,a cross-regional deployment strategy is derived using an adaptive genetic algorithm.The proposed algorithm is evaluated using a real-world dataset from Chengdu,Sichuan Province.Experimental results demonstrate the effectiveness of the proposed solution.

Tropical Cyclone Forecasting Based on Cross-modal Fusion and Multi-generators

LIU Qian, SUN Hu, GUI Yaocheng, ZHOU Guoqiang

Computer Science. 2025, 52 (12): 81-91. doi:10.11896/jsjkx.250100030

Abstract

PDF(3632KB) ( 15 )

References | Related Articles | Metrics

Accurately predicting the trajectory and intensity of tropical cyclones(TCs) is essential for disaster mitigation and prevention.Although deep learning-based advances have demonstrated great prediction performance,a majority of these methods only focus on unimodal data and overlook the rich correlations between different modalities.To fully leverage the rich information contained in multimodal data,a novel framework for tropical cyclone forecasting based on cross-modal fusion and multi-generators is proposed.The framework includes a multimodal feature extraction module,a cross-model feature fusion module and a generative adversarial network(GAN).The multimodal feature extraction module obtains feature representations from the best historical trajectory data,atmospheric reanalysis data and environmental field data respectively.The cross-model feature fusion module fuses multimodal features through a novel cross-modal feature complementation strategy.GAN produces final TC predictions through multiple generators.Additionally,a feature fusion loss is constructed to help boosting the model’s performance.Experiments show that the proposed method can not only maintain high efficiency in training and inference stages,but also achieve better prediction performance.

Identifying Influential Nodes in Multilayer Networks Based on Layer Weighting and Gravity Centrality

WANG Jianbo, LUO Yu, XU Xiaoke, DU Zhanwei, LI Ping

Computer Science. 2025, 52 (12): 92-101. doi:10.11896/jsjkx.241000090

Abstract

PDF(4493KB) ( 11 )

References | Related Articles | Metrics

Identifying key nodes in multilayer networks is a major research focus in network science,as it plays a crucial role in understanding network structure and function.Inspired by the gravity model,most existing methods primarily rely on local or global topological information,often overlook the influence of intra-layer and inter-layer structures on nodes in multilayer networks.This oversight limits the effectiveness of node identification.To address this,this paper introduces a layer weighting and gravity centrality algorithm for identifying key nodes in multilayer networks.The algorithm first assigns weights to each network layer by considering both intra-layer and inter-layer structures,thus quantifying the influence of degree centrality across different layers.Next,it incorporates the impact of inter-layer structures on propagation paths to define the effective distance between nodes.Finally,the influence of each node within the entire network is calculated using a gravity-based formula.Extensive experiments on nine real-world networks show that the proposed algorithm offers higher accuracy and resolution compared to six benchmark methods.

Protein Complex Identification Algorithm Based on Hypergraph Network Embedding

WANG Jie, YANG Xiancan, ZHAO Xingwang

Computer Science. 2025, 52 (12): 102-114. doi:10.11896/jsjkx.250900062

Abstract

PDF(8620KB) ( 9 )

References | Related Articles | Metrics

Protein complexes are crucial for understanding cellular functions and identifying biological processes,playing critical roles in cell biology.The use of network clustering in PPI networks to identify protein complexes has become a hot research topic in data mining and bioinformatics.A variety of computational methods have emerged to identify protein complexes.However,most existing algorithms primarily use original network to detect dense subnetworks and fail to break through the limitations of traditional graph structures for multi-node interactions.Aiming at the issue of many-to-many complex interaction characteristics prevalent in biological networks,this paper proposes a novel protein complex identification method based on hypergraph network embedding(PCIHNE).Through the ability of hypergraph networks,it firstly converts the original PPI network into a hypergraph network.Then a hierarchical compression strategy recursively compresses the hypergraph into multiple smaller hypergraphs at different levels,thereby constructing a multi-scale analysis framework.Next,hypergraph convolution is performed on each levels to generate node representations at different granularities.These node representations are concatenated to obtain the complete node representation.Based on the representations obtained from hypergraph learning,a weighted PPI network is constructed by similarity on the original network.Finally,a core-attachment based strategy is used to obtain predicted protein complexes in the weighted PPI network.It evaluates the effectiveness of PCIHNE by comparing it with other protein complex algorithms on multiple yeast and human datasets.Experimental results demonstrate that PCIHNE is better than comparison protein complex identification methods regarding F-measure and Accuracy metrics.

PIEnum:Efficient Algorithm of Path Enumeration on Large Uncertain Graphs

XIE Wenlin, DU Ming, ZHOU Junfeng

Computer Science. 2025, 52 (12): 115-124. doi:10.11896/jsjkx.241100090

Abstract

PDF(4363KB) ( 9 )

References | Related Articles | Metrics

A basic approach of investigating the relationship between two vertices on the uncertain graph is to enumerate all paths between them.To solve the problem of inadequate pruning and redundant computation in the state-of-the-art algorithms,this paper proposes an pruning and index based algorithm,PIEnum,whose target is to enumerate simple paths from a source vertex s to a target vertex t where the length of each path is no more than a given hop constraint k and the probability of each path is no less than a given probability threshold γ.For an input query,it firstly excludes the unpromising vertices to reduce the search space.Then it builds an online light-weight index to avoid repeated pruning examinations during the enumeration.Finally,it develops an efficient approach to prune invalid search branches during the enumeration.To further improve the performance on dense graphs,it implements PIEnum⁺ based on the Join paradigm.The comprehensive experimental results on 10 real-life graphs show that PIEnum improves the overall performance by at least 10 times compared to the state-of-the-art algorithms.

Locality-aware Cache Management Strategy for Concurrent Graph Analysis

LI Hanqiao, ZHAO Yuanjun

Computer Science. 2025, 52 (12): 125-132. doi:10.11896/jsjkx.250200062

Abstract

PDF(4195KB) ( 7 )

References | Related Articles | Metrics

With the rapid growth of graph computing,modern graph platforms routinely execute a large number of concurrent graph analytics tasks to extract the latent value in massive datasets.Consequently,concurrent graph processing has been widely adopted in domains,including intelligent education,public administration,and news media.However,most existing graph proces-sing systems are originally designed for single-task execution and suffer from excessive redundant data accesses when handling concurrent workloads.Although prior studies have observed significant redundancy in in-memory graph data across concurrent tasks and have attempted to exploit temporal and spatial locality to share underlying graph data,they largely overlook the data locality in private state updates.This limitation leads to low cache utilization and,ultimately,degraded system throughput.To address this challenge,this paper proposes CCG,a locality-aware cache management strategy for concurrent graph analysis,which fully exploits both temporal and spatial locality across tasks to reduce redundant data accesses and synchronization overhead.Specifically,CCG efficiently buffers and incrementally merges redundant updates,leveraging data locality to perform high-throughput batch updates in memory.This design minimizes access costs,mitigates cache thrashing,and significantly improves concurrency performance.Moreover,CCG employs a multi-level cache hierarchy to enable layered buffering and merging,thereby eliminating synchronization and locking overhead during private state updates.Experimental results show that CCG improves system throughput by 2.3×～7.8× over GRASP.

SPP-STGCN:Spatio-Temporal Graph Convolutional Network for Pedestrian Trajectory Predictionwith Scene-Perdestrian-Perdestrain Interactions

HONG Mingjun, JI Qingge

Computer Science. 2025, 52 (12): 133-140. doi:10.11896/jsjkx.241200212

Abstract

PDF(2187KB) ( 11 )

References | Related Articles | Metrics

Pedestrian trajectory prediction is a fundamental and critical task in autonomous driving and intelligent surveillance systems.The constraints of the scene are one of the important factors affecting pedestrian movement trajectories.Despite existing research efforts to incorporate scene factors into trajectory prediction,these methods often fall short in integrating scene information,particularly in terms of comprehensive scene fusion.To overcome these limitations,this study proposes a new pedestrian trajectory prediction model,namely SPP-STGCN.The SPP-STGCN model adopts a two-stage architecture to enhance prediction accuracy.In the first stage,the model integrates trajectory and scene data.Through the Scene Adjacency Fusion Block(SAFB),the model fuses these two types of features to construct a spatio-temporal graph adjacency matrix that incorporates scene features,thereby providing rich contextual information for prediction.Concurrently,the model operates in parallel along the temporal and spatial dimensions,constructing pedestrian trajectory spatio-temporal graphs that describe temporal and spatial correlations based on trajectory information.In the second stage,scene-graph convolutional networks extract features from the temporal and spatial spatio-temporal graphs.The extracted features are then fused and processed through a temporal pyramid extrapolation convolution to obtain the two-dimensional Gaussian distribution of the pedestrian’s future trajectory.Finally,SPP-STGCN uses this distribution as a probabilistic model for predicting pedestrian trajectories,generating future trajectories through sampling.Comparative experimental results on the ETH and UCY public datasets show that the SPP-STGCN model has achieved the current state-of-the-art performance in comparison experiments with nine mainstream models.Ablation experiments and qualitative analysis further confirm the effectiveness and rationality of the proposed model.The SPP-STGCN model significantly enhances pedestrian trajectory prediction performance by effectively integrating scene features.

Appearance Enhancement and Semantic Segmentation-based Neural Radiance Fields

CAO Mingwei, HUANG Baolong, ZHAO Haifeng

Computer Science. 2025, 52 (12): 141-149. doi:10.11896/jsjkx.250400075

Abstract

PDF(3683KB) ( 13 )

References | Related Articles | Metrics

The accelerated advancement of deep learning has notably propelled 3D reconstruction techniques within the field of computer vision.NeRFs have become an essential methodology due to their adeptness at scene modeling and superior view synthesis.However,challenges persist in dynamic environments,particularly in managing intricate lighting variations and transient object interference.Alterations in imaging conditions may lead to inconsistent scene appearances,thereby degrading the quality of view synthesis.Concurrently,dynamic elements can adversely affect the photorealism of reconstructed scenes.To mitigate these issues,this paper introduces an AS-NeRF.By incorporating frequency regularization and composite positional encoding into the sampling strategy,AS-NeRF enhances the efficiency of appearance feature fusion,thereby augmenting the model’s adaptability to variations in lighting and camera parameters.This subsequently improves color consistency and overall rendering realism.Additionally,a lightweight segmentation network is utilized to predict transient visibility masks in an end-to-end manner,effectively isolating dynamic objects and reducing their impact on view synthesis quality.The efficacy of AS-NeRF is verified through experiments conducted on the Photo Tourism datasets,which are compared qualitatively and quantitatively with several existingme-thods.The experimental results demonstrate that AS-NeRF surpasses existing approaches in terms of synthesis accuracy and further confirms the accuracy of the computed segmentation masks in distinguishing transient objects.

ETF-YOLO11n:Object Detection Method Based on Multi-scale Feature Fusion for TrafficImages

XIA Shufang, YIN Haonan, QU Zhong

Computer Science. 2025, 52 (12): 150-157. doi:10.11896/jsjkx.241200021

Abstract

PDF(5245KB) ( 14 )

References | Related Articles | Metrics

Deep learning algorithms have made significant progress in the field of computer vision in recent years,but the accuracy of object detection in complex traffic scenes is still unsatisfactory due to the small size of traffic objects,inconspicuous feature,and susceptibility to interference.To address this problem,this paper improves the state-of-the-art YOLO11 and designs the ETF-YOLO11n based on multi-scale feature fusion.Firstly,it designs TFF,which effectively fuses the feature information of different sizes extracted from the backbone.Secondly,it designs HDCFE,effectively integrates the features extracted from different receptive fields and reduces the interference on the detection effect of the model due to occlusion and overlapping.Finally,the proposed GeoCIoU is used to replace CIoU,and the model can provide more accurate feedback on the matching of the predicted box and the ground-truth box through the two different penalization terms.The ETF-YOLO11n achieves an AP of 65.6% and mAP@0.5 of 90.7% on KITTI dataset,which is improved by 2.4 percentage points and 1.2 percentage points.In addition,ETF-YOLO11n achieves 42.5% AP and 59.8% mAP@0.5 on COCO-Traffic,and EFT-YOLOv8n achieves 66.9% AP and 91.5% mAP@0.5 on KITTI dataset.The results show that the proposed methods significantly improve the performance and have good ge-neralization ability to different models and datasets,achieve a good balance between the accuracy and parameters.The source code has been opened.

Ancient Mural Image Restoration Network Using Involution Cascaded Attention Mechanism

ZHOU Qixue, YU Ying, HU Jialv

Computer Science. 2025, 52 (12): 158-165. doi:10.11896/jsjkx.241000124

Abstract

PDF(4974KB) ( 11 )

References | Related Articles | Metrics

Chinese ancient murals are precious cultural heritage of humanity,recording the social,religious,cultural,and artistic activities of people in various regions of China throughout history.Due to prolonged exposure to the natural environment,many murals have developed defects such as cracks,scratches,corrosion,and even large-scale peeling.Therefore,the protection and restoration of murals are urgently needed.The digital restoration technology for damaged murals has become an important means to solve this problem by reconstructing the structure and texture of the mural images and virtually filling the damaged areas.Most existing mural image restoration methods are hard to effectively restore missing mural content with complex structures and rich color variations.In response to this issue,this paper proposes an ancient mural image restoration network using the involution cascade attention mechanism.The network firstly uses involution operations instead of traditional convolutions to improve the quality of repairing damaged mural textures and colors.Secondly,a cascaded attention module is proposed to capture contextual information at different scales,which can better repair damaged areas of murals of different sizes.Thirdly,FFC residual blocks are introduced to capture global structural information to enhance the network’s color restoration ability for damaged areas of murals.This article conducts experiments on simulated and real damaged mural datasets,comparing the restoration results with four other classic methods.The experimental results show that the proposed model outperforms other comparative methods in restoring the clarity,color consistency,and structural continuity of mural textures.

Railway Fastener Segmentation Method Based on Sc-DeepLabV3+ Model

HUANG Kun, HE Lang, WANG Zhanqing

Computer Science. 2025, 52 (12): 166-174. doi:10.11896/jsjkx.241000130

Abstract

PDF(4262KB) ( 10 )

References | Related Articles | Metrics

The deterioration of track fasteners is a critical factor affecting railway traffic safety.Utilizing deep learning image re-cognition methods for segmenting images collected by track fastener detection robots can significantly improve the efficiency of fastener defect detection.This paper addresses the current lack of publicly available datasets for track fasteners and the challenges posed by complex backgrounds that increase segmentation difficulty and processing time.This paper manually creats the RFS(Rail Fastener Segmentation) track fastener dataset and proposes a segmentation method based on the Sc-DeepLabV3+ model.By replacing the backbone network of the original DeepLabV3+ model with the lightweight MobileNetV4,it accelerates computation speed and introduces an improved S-ASPP(Switchable Atrous Spatial Pyramid Pooling) module to enable the network to achieve denser pixel sampling,enhancing its ability to extract detailed features.Additionally,it incorporates the CSWin(Cross-Shaped Window Self-Attention) attention mechanism to compute horizontal and vertical attention in parallel,reducing interference from complex backgrounds.In the experimental section,this paper proposes the RailAugment data augmentation technique to effectively increase the diversity and coverage of the dataset,ultimately resulting in a total of 6 832 images,including 4 782 for training,1 366 for validation,and 684 for testing.Experimental results show that the mIoU and mPA reach 95.17% and 97.14%,respectively,which represent improvements of 2.19 percentage point and 0.36 percentage point compared to the original model.Although the performance improvement is relatively small,significant improvements are observed in detailed feature extraction and background interference handling.Furthermore,the Sc-DeepLabV3+ model is validated on the DeepGlobe dataset,demonstrating its robustness and generalization ability.Its inference speed is 51.4 ms and 66.5 ms faster than the mainstream Swin-UNet and Segmenter models,respectively,showing good efficiency and real-time performance.Therefore,this model has broad application potential in railway maintenance and other fields,effectively reducing labor and computational costs while improving detection efficiency.

Road Crack Detection Method for Embedded Applications

HU Peng, XIA Xiaohua, ZHONG Yuquan

Computer Science. 2025, 52 (12): 175-188. doi:10.11896/jsjkx.241200214

Abstract

PDF(11836KB) ( 10 )

References | Related Articles | Metrics

Aiming at the problems of complex models and slow processing speed of pavement crack detection model based on deep learning in embedded platform deployment and application,a pavement crack detection method for embedded application is proposed based on YOLO.Firstly,the two-stage concatenated convolutional module is constructed in the backbone network to optimize the feature perception of the convolution channel and the input space,and the Smooth Intersection over Union(SIoU) considering the vector angle between the prediction bounding boxes and ground truth bounding boxes is used as the loss function of the network to improve the accuracy and speed of the prediction bounding boxes regression.Secondly,a method of the cascade channel pruning and convolution-by-convolution weight pruning is proposed,and channel pruning and weight pruning are performed successively,the unnecessary channels are removed and the redundant weights are removed by convolution,which significantly compresses the model while ensuring the stability of the model.Then,distill the generalization ability of the YOLOv5l model for crack features into the pruned model to further improve its ability to characterize cracks.Finally,under the TensorRT engine,the inference speed of the model is improved through layer optimization and other methods.The experimental results show that compared with the original model,the mean Average Precsion of the proposed method is increased by 2.7%,the model parameters and model volume are reduced by 78.24% and 76.13% respectively,and the detection rate is increased by 587.88 %.The model is deployed on NVIDIA Jetson Nano embedded devices through TensorRT for testing,the detection accuracy is unaffected,and the detection rate is increased by 140%,compared with the commonly used models such as YOLOv5-Lite and the la-test models such as YOLOv10 and RT-DERT,it has the highest detection accuracy and the most significant lightweight effect,which is suitable for embedded applications.

Superpixel-level Graph Feature Learning Method for Hyperspectral Image Denoising

WU Ying, YE Hailiang, CAO Feilong

Computer Science. 2025, 52 (12): 189-199. doi:10.11896/jsjkx.250100082

Abstract

PDF(5942KB) ( 9 )

References | Related Articles | Metrics

Hyperspectral image denoising methods based on traditional deep learning usually have difficulty capturing the long-range correlation of spatial positions and the similarity of global irregular local blocks,resulting in loss of detailed information and insufficient structural integrity after denoising.To this end,this paper proposes a new superpixel-level graph feature learning method for hyperspectral image denoising,which aims to use graph neural networks to extract spatial-spectral features and capture the long-range correlation of spatial positions of irregular local blocks to preserve texture details and structural information.Firstly,a gated attention module is designed to suppress noise and enhance spectral correlation,laying the foundation for subsequent superpixel segmentation.Then,a superpixel-level graph aggregation module is designed,which effectively maintains the structural integrity and clarity of internal details of the hyperspectral image by segmenting the hyperspectral image into multiple spatially connected superpixels according to the spatial dimensions and using a shared linear layer to learn the weighted values of pixels in the superpixel.Then,graph convolution is used to update the node information.Finally,nuclear norm regularization is introduced in the training process for constraint by taking advantage of the low rank of hyperspectral images,and a low-rank-spatial spectral denoising loss is proposed to focus on preserving structural information.Experimental results show that the proposed method outperforms the current advanced methods in performance.

Structures of Multi-line Standard Contradictions in First-order Logic

ZENG Dan, HE Xingxing, LI Yingfang, LI Tianrui

Computer Science. 2025, 52 (12): 200-208. doi:10.11896/jsjkx.250200060

Abstract

PDF(1537KB) ( 14 )

References | Related Articles | Metrics

Automatic reasoning constitutes a critical research domain in artificial intelligence,where inference rules play a pivotal role in determining its efficiency.Deductive reasoning based on the separation of contradictions represents a reliable and complete inference rule,offering advantages such as multi-literal collaboration,multi-clause integration,and dynamic deduction.The construction method of contradictions significantly impacts the efficiency of deduction.In light of this,this paper first introduces a type of contradiction structure grounded in propositional logic,i.e.,multi-line standard contradictions,and elaborates on the compound strategies for such contradictions under various scenarios.Specifically,it outlines the conditions and methods for generating new standard contradictions by leveraging two multi-line standard contradictions,while distinguishing between common and characteristic conclusions regarding their compound properties.Secondly,the paper highlights that incorporating specific literals into clauses can transform multi-line standard contradictions into more streamlined linear standard contradictions,providing lite-rals addition strategies tailored to such contradictions.Additionally,an algorithm for generating multi-line standard contradictions within propositional logic is designed.Finally,the structure and properties of double-line contradictions and complete standard contradictions based on first-order logic are presented.

Complexities of Logic Programs with Convex Aggregates

WANG Xianglong, WANG Yisong, XIE Zhongtao

Computer Science. 2025, 52 (12): 209-214. doi:10.11896/jsjkx.241000148

Abstract

PDF(1446KB) ( 9 )

References | Related Articles | Metrics

ASP is a paradigm for descriptive problem solving and is widely used in fields such as planning,diagnosis,scheduling and bioinformatics.To enhance the expressive power of ASP,it has introduced aggregate functions from database systems and proposed semantics such as FLP and SPT.The specific form of abstract constraint stripping aggregation function constraints has become an important tool for studying the semantics and other properties of ASP,and has yielded relevant results regarding the relationships between various answer set semantics of abstract constraint logic programs and complexity issues.The article further investigates the properties of abstract constraint logic with only convex abstract constraint atoms,proving that the existence of FLP answer sets for regular logic programs containing only convex abstract constraint atoms is Σ^p₂-complete,with its cautious reasoning and brave reasoning being Π^p₂-complete and Σ^p₂-complete,respectively.These complexity results further clarify the expressive power relationships between various types of logic programs,providing new ideas for the design of effective answer set solvers and laying a theoretical foundation for further exploration of ASP applications in solving problems represented by convex abstract constraints.

Unsupervised Dialogue Topic Segmentation Method Based on Utterance Rewriting

LI Tongliang, LI Qifeng, HOU Xia, CHEN Xiaoming, LI Zhoujun

Computer Science. 2025, 52 (12): 215-223. doi:10.11896/jsjkx.241000136

Abstract

PDF(2161KB) ( 7 )

References | Related Articles | Metrics

Dialogue Topic Segmentation(DTS) task aims to automatically divide a multi-turn conversation into different topic segments,enabling more precise understanding and processing of dialogue content.DTS plays an important role in dialogue modeling tasks.Traditional DTS methods primarily rely on semantic similarity and dialogue coherence to perform unsupervised topic segmentation,but these features are often insufficient to fully capture complex topic transitions in conversations,and unannotated dia-logue data has not been fully explored and utilized.To address this issue,recent DTS methods employ adjacent utterance ma-tching and pseudo-segmentation to learn topic-aware representations from dialogue data,further extracting useful cues from unannotated dialogues.However,common phenomena such as coreference and ellipsis in multi-turn dialogues may affect the calculation of semantic similarity,thereby weakening the accuracy of adjacent utterance matching.To solve this problem and fully leverage the useful cues in dialogue relationships,this study proposes a novel unsupervised DTS method that combines utterance rewriting(UR) techniques with unsupervised learning algorithms.This approach rewrites coreferential and elliptical expressions in the dialogue to restore them to their complete forms,better capturing the thematic cues in the conversation.Experimental results show that the proposed utterance rewriting topic segmentation model(UR-DTS) significantly improves topic segmentation accuracy,achieving state-of-the-art performance.On the DialSeg711 dataset,the error rate P_k and WinDiff(WD) improves by approximately 6 percentage point,reaching 11.42% and 12.97%,respectively.On the more complex Doc2Dial dataset,P_k and WD improve by 3 percentage point and 2 percentage point,reaching 35.17% and 38.49%.These results demonstrate that UR-DTS has a significant advantage in capturing topic transitions in conversations and shows greater potential for leveraging unannotated dialogue data.

Adverse Drug Reaction Relationship Extraction Based on Chain of Thought Enhancement UnderHigh and Low Resources

LI Hao, YANG Yumeng, ZHAO Boyang, ZHENG Puqi, LIN Hongfei

Computer Science. 2025, 52 (12): 224-230. doi:10.11896/jsjkx.250600140

Abstract

PDF(3410KB) ( 7 )

References | Related Articles | Metrics

ADRs refer to harmful and unintended responses that occur when drugs are administered at normal doses for prevention,diagnosis,treatment of diseases,or regulation of physiological functions.These reactions are unexpected and can significantly impact patients’ health and physical condition.To identify ADR-related relational triples from social media and biomedical literature data—providing early warnings for patients and healthcare systems,as well as references for pharmaceutical research and drug development－this paper proposes two frameworks based on generative large language models for both low-resource and high-resource annotated data scenarios:the ADR-GPT framework,which extracts ADR relations using Chain-of-Thought promp-ting,and the CADRE-LLM framework,which leverages Chain-of-Thought-enhanced fine-tuning.It evaluates the performance of these two frameworks on the Twitter and PubMed datasets.CADRE-LLM achieves F1 score improvements of 1.51 percentage points and 1.74 percentage points over previous state-of-the-art models on the respective datasets.The low-resource ADR-GPT framework also demonstrates strong performance,outperforming fully supervised fine-tuned Qwen2.5 on the PubMed dataset.Ablation studies further validate the effectiveness of each module within the two proposed frameworks.

MemLong:Memory-augmented Retrieval for Long Text Modeling

LIU Weijie, TANG Zecheng, LI Juntao

Computer Science. 2025, 52 (12): 231-238. doi:10.11896/jsjkx.250100094

Abstract

PDF(3735KB) ( 8 )

References | Related Articles | Metrics

Recent advancements in Large Language Models (LLMs) have yielded remarkable success across diverse fields.However,handling long contexts remains a significant challenge for LLMs due to the quadratic time and space complexity of attention mechanisms and the growing memory consumption of the key-value cache during generation.To address this issue,this paper proposes MemLong’a memory-augmented method for long-text modeling,which enhances long-context language modeling by leveraging an external retriever to access historical information.MemLong integrates a non-parametric retrieval-memory module with a partially trainable large language model,and introduces a fine-grained,controllable retrieval attention mechanism that effectively utilizes semantically relevant text blocks.The non-parametric module is responsible for retrieving relevant historical information from an external knowledge base,while the LLM generates outputs by fusing this retrieved information with the current input.The proposed attention mechanism allows the model to dynamically adjust its focus on the retrieved information during generation.Comprehensive evaluations on multiple long-context language modeling benchmarks demonstrate that MemLong consistently outperforms other state-of-the-art LLMs.Furthermore,MemLong significantly enhances the model’s capacity to process long texts.On a single NVIDIA 3090 GPU,MemLong can scale the effective context length from 4 000 to 80 000 tokens,representing a 20-fold increase.This breakthrough enables MemLong to process longer input texts,leading to a better understanding and ge-neration of long-form content.It provides new possibilities for tackling ultra-long text tasks and opens up promising new directions for future research in long-text language modeling.

Line of Sight Guided Self Expert Cloning with Reinforcement Learning for Unmanned SurfaceVehicle Path Tracking

LIU Jiahui, ZHAO Yinuo, TIAN Feng, QI Guangpeng, LI Jiangtao, LIU Chi

Computer Science. 2025, 52 (12): 239-251. doi:10.11896/jsjkx.250200059

Abstract

PDF(4348KB) ( 9 )

References | Related Articles | Metrics

Unmanned Surface Vehicle(USV) path tracking is crucial for marine autonomous operations,as environmental factors such as wind,waves,currents,and USV’s control errors can affect tracking performance.Reinforcement learning(RL),with its online interaction and real-time feedback,offers a promising approach for actively adapting to dynamic environments.However,its trial-and-error training process poses safety risks in real-world applications,and the gap between ideal simulation environments and complex real-world conditions further limits its practical effectiveness.To address these challenges,this paper proposes LECUP(Line-of-sight-guided self-Expert Cloning for USV Path tracking),a new algorithm designed for complex marine environments.LECUP first trains an RL expert in a still water environment and then uses self-expert cloning to transfer the agent to a more complex environment.To ensure effective knowledge transfer,LECUP introduces a data filling mechanism,where the ex-periences accumulated by the self-expert in the still-water environment are dimensionally padded and stored for initializing the agent in the complex environment.Then,reinforcement learning is used to fine-tune the agent in the complex environment,further enabling adaptation to the complexities of the environment.Moreover,LECUP incorporates a line-of-sight guidance module to calculate the target heading,decoupling the path tracking control from the specific geometry of the path and enhancing the USV’s adaptability to various path shapes.This method enables ongoing policy refinement in complex environments while mitigating safety risks associated with random initialization.Extensive experimental results show that LECUP performs better than baseline methods in path tracking tasks,especially under challenging conditions.

Data and Knowledge Enhanced Medical Visual Question Answer Network

YAN Yujing, HOU Xia, GUO Yuting, ZHANG Mingliang, SONG Wenfeng

Computer Science. 2025, 52 (12): 252-259. doi:10.11896/jsjkx.241000105

Abstract

PDF(3305KB) ( 7 )

References | Related Articles | Metrics

Med-VQA aims to accurately answer clinical questions based on a given medical image,which is key in advancing clinical medical intelligence.Despite some progress in this field,challenges remain in extracting deep multimodal information from both images and questions and in effectively training models on small-scale datasets.To address these issues,this paper proposes a Med-VQA network that incorporates dual data and knowledge enhancement.Aiming at small-scale datasets,a multimodal conditional mixing module is designed to enhance the input image and question data,and linear combinations of input sample pairs are performed by using the category of questions as constraints to improve the rationality of answer generation.For multimodal feature extraction,an image location recognizer based on convolutional neural networks is designed to encode the captured image location features into the fusion process of image and question features for knowledge enhancement,which can effectively achieve feature extraction under fewer parameters.Experimental results on the SLAKE and VQA-RAD datasets demonstrate that the proposed model significantly outperforms the baseline models.

Explanation Method for Entity Alignment Based on Neighborhood Matching Probability andType Quotient Graph

ZHANG Xiaoming, QIU Jingjing, WANG Huiyong

Computer Science. 2025, 52 (12): 260-270. doi:10.11896/jsjkx.241100081

Abstract

PDF(4550KB) ( 5 )

References | Related Articles | Metrics

In recent years,a variety of embedding-based entity alignment methods have emerged,which achieve entity alignment by mapping entities and relations into low-dimensional vector spaces and calculating these vector representations.Although these methods have achieved good performance,there are relatively few studies on their explainability.Therefore,a post hoc explanation algorithm for entity alignment(PE-EA) is proposed to generate explanations for the prediction results of the entity alignment model based on knowledge graph embedding.The method first evaluates the functionality of entities and their relations by calculating their connection counts in the knowledge graph,and then quantifies the importance of the entity neighborhood structure.Subsequently,the entity functionality is combined with the embedding vector to calculate the embedding vector of the relation,and the matching probability of the relation pair is obtained based on it.Then,based on the neighborhood information of the predicted entity pair,the matching probability of the candidate explanation pairs in the neighborhood is calculated,and the explanation triplets of the predicted entity pair are screened out and combined into an explanation subgraph.Furthermore,the concept of type quotient graph is introduced to abstract the explanatory subgraph,compress the data and simplify the explanation generation process,thereby reducing the number of candidate explanations while enhancing the quality and effectiveness of the explanations.On five commonly used experimental datasets,the two evaluation indicators of fidelity and sparsity are used to verify that the explanations generated by the proposed model have high accuracy and simplicity.

Survey of Data Prediction Methods Using Artificial Intelligence in the Financial Sector

CHEN Xiayi

Computer Science. 2025, 52 (12): 271-284. doi:10.11896/jsjkx.250700166

Abstract

PDF(3289KB) ( 10 )

References | Related Articles | Metrics

Given the high complexity of financial markets and the inherently noisy,non-linear nature of their data,artificial intelligence(AI),particularly machine learning and deep learning,has emerged as a core driving force in financial data prediction.This paper systematically summarizes the latest research progress and core methodologies in this domain over the past three years.In traditional machine learning,the trend has shifted from applying single learners toward sophisticated model fusion strategies,such as Stacking,and hybrid paradigms that integrate optimization algorithms for automated feature selection and hyperparameter tu-ning.Deep learning applications demonstrate a clear evolutionary trajectory,starting with foundational Recurrent Neural Networks(RNNs) enhanced by data decomposition and attention mechanisms,and progressing to hybrid architectures like CNN-RNN for capturing multi-dimensional features.This paper further details the adoption of cutting-edge models:Graph Neural Networks(GNNs) for modeling entity relationships,Transformers for capturing long-range dependencies,and Deep Reinforcement Lear-ning(DRL) for shifting the paradigm from prediction to autonomous decision-making.Furthermore,the review outlines key feature engineering techniques and introduces an innovative “structured neural network modeling” paradigm,which proposes aligning the model’s architecture with the intrinsic topology of financial products to enhance interpretability.Finally,this paper synthesizes the core challenges facing the field—including data quality,model robustness,and interpretability－and provides a forward-looking perspective on future research directions such as deep multi-modal fusion,causal inference,financial Large Language Models(LLMs),and explainable AI(XAI).

Energy-efficient Trajectory and Resource Optimization for Multi-cluster NOMA-UAV Networks

LI Zhike, XU Wanping

Computer Science. 2025, 52 (12): 285-293. doi:10.11896/jsjkx.250100016

Abstract

PDF(3170KB) ( 8 )

References | Related Articles | Metrics

An optimization scheme is proposed to ensure QoS guarantee in UAV-assisted multi-cluster NOMA downlink networks with limited resources.In this paper,the UAV serves as an airborne mobile base station to communicate with ground users.Due to limited energy,hovering time is introduced as an optimization variable to allocate more energy for communication.The total throughput is maximized by optimizing user clustering,intra-cluster power allocation,and communication time allocation.Due to its non-convexity,the optimization problem is divided into three sub-problems.The power allocation problem is addressed using the SCA method,and the communication time allocation is solved via linear programming.First of all,the Mean Shift algorithm is employed for user clustering.Unlike K-means,it clusters users by calculating local density peaks,ensuring higher intra-cluster user concentration.Then,an improved Mean Shift algorithm is proposed to balance user distribution by splitting oversized clusters,thereby ensuring individual user QoS.Finally,an original cluster head hovering scheme is introduced to avoid increasing the UAV’s flight distance due to additional sub-clusters,then GA is used for trajectory optimization,enhancing total throughput by reducing the UAV’s non-communication energy consumption while ensuring user QoS.The optimization scheme has low computational complexity and strong real-time performance.Simulation results show that the optimization scheme with an improved Mean Shift algorithm reduces the non-communication energy consumption than the K-means algorithm,and improves the system throughput by an average of 5.94% at different transmit power and energy efficiency by an average of 6.82% at different number of users.

Efficient Clustering Routing Method for WSNs Based on Clone Reverse Learning Grey WolfOptimization Algorithm

CHEN Haiyan

Computer Science. 2025, 52 (12): 294-301. doi:10.11896/jsjkx.250200116

Abstract

PDF(2218KB) ( 9 )

References | Related Articles | Metrics

To address the issues of uneven node energy consumption and optimal cluster head selection in clustering routing for WSNs,this paper proposes a Clone Reverse Learning Grey Wolf Optimizer-based Energy-Balanced Routing Protocol(CRLGWORP).This algorithm introduces a clone selection mechanism into the traditional grey wolf optimizer framework,enhancing population diversity by replicating high-quality individuals,and combines reverse learning strategies to expand the search space for solutions,effectively improving global optimization capabilities.An adaptive weighting function is designed with the objectives of maximizing the network’s average residual energy and minimizing the average distance from cluster heads to the base station.The weights are dynamically adjusted based on the network’s energy distribution to balance the optimization focus between energy efficiency and communication distance.In the cluster head election phase,nodes with high energy and proximity to the base station are prioritized.During the data transmission phase,a multi-hop gradient relay mechanism is employed to optimize communication paths,reducing energy consumption for long-distance transmissions.Experimental results demonstrate that,compared with LEACH,LEACH-C,HEED,FIGWO and HGWCSOA-OCHS algorithms,the proposed algorithm significantly extends the network lifespan and improves node energy balance.

Research on Foundation Model Methods for Addressing Non-IID Issues in Federated Learning

WANG Xin, CHEN Kun, SUN Lingyun

Computer Science. 2025, 52 (12): 302-313. doi:10.11896/jsjkx.241200056

Abstract

PDF(2313KB) ( 7 )

References | Related Articles | Metrics

Federated learning,due to its inherent privacy-preserving nature,has gradually become a widely recognized framework for distributed machine learning.However,it faces significant challenges such as insufficient generalization performance,degraded convergence efficiency,and data skew,particularly in the presence of Non-IID data.Using pre-trained foundation models to mitigate Non-IID issues has emerged as a novel approach,leading to the development of various solutions.This review examines exis-ting works from the perspective of pre-trained foundation models.Firstly,it introduces foundation model methodologies and provides a comparative analysis of typical foundation model architectures.Secondly,a new classification framework is proposed from three perspectives:modifying inputs,retraining parts of the foundation model,and parameter-efficient fine-tuning.Finally,it explores the core challenges of this type of work and outlines future research directions.

Decentralized Federated Learning Algorithm Sensitive to Delay

PENG Jiao, CHANG Yongjuan, YAN Tao, YOU Zhangzheng, SONG Meina, ZHU Yifan, ZHANG Pengfei, HE Yue, ZHANG Bo, OU Zhonghong

Computer Science. 2025, 52 (12): 314-320. doi:10.11896/jsjkx.241100085

Abstract

PDF(2197KB) ( 9 )

References | Related Articles | Metrics

In recent years,the rapid development of deep learning,mobile devices,and IoT technology has led to a surge in demand for model inference and data storage on edge devices.Traditional centralized model training methods are limited by datavo-lume,communication bandwidth,and user data privacy issues and cannot effectively address the new challenges.Therefore,federated learning technology is born.Federated learning allows edge devices to train models based on local data and upload model parameters to a central server for aggregation and distribution,ensuring that joint modeling can be performed without data leaving the trusted domain of each party.Furthermore,distributed federated learning has been developed to overcome issues such as latency,bandwidth limitations,and single point of failure risks.However,the training efficiency of federated learning is severely affected by real-world network delay and bandwidth factors,making multi-party joint modeling difficult.To address this issue,this paper proposes a decentralized federated learning algorithm DBFedAvg that dynamically selects nodes with lower average delay as the main nodes to reduce communication costs and improve global model training performance,accelerating model convergence.Experimental results on the Sprint network and other scenarios have validated that the proposed method brings significant improvements in communication costs and model convergence.

Malware Detection Based on API Sequence Feature Engineering and Feature Learning

YANG Yizhe, LU Tianliang, PENG Shufan, LI Xiaolin

Computer Science. 2025, 52 (12): 321-330. doi:10.11896/jsjkx.250300056

Abstract

PDF(2502KB) ( 7 )

References | Related Articles | Metrics

API sequence-based malware analysis methods can effectively capture the behavioral characteristics of programs during runtime.However,existing detection approaches typically focus solely on API names while neglecting parameters and return values,or fail to adequately explore their semantic information and inter-parameter correlations,resulting in limited detection performance.To address this,this paper proposes a malware detection method combining systematic feature engineering with a deep neural network architecture.Specifically,the method implements structured encoding of API sequences based on the data characteristics of API names,parameters,and return values.Multiple RefConv convolutional blocks are then employed to extract multi-scale features for each API call.Finally,the feature vectors are fed into a parallel recurrent neural network based on BiGRU-BiLSTM to learn long-term and short-term dependencies within API sequences.Experiments conduct on a dataset containing 25 000 API sequences,this method achieves 93.55% accuracy in comprehensive performance tests.Validation through temporal concept drift,spatial concept drift,and ablation experiments demonstrates that the proposed method can effectively detect malware.

Lightweight Secure Authentication and Key Update Scheme for 5G Urban Transportation

SU Xinzhong, XU Youyun

Computer Science. 2025, 52 (12): 331-338. doi:10.11896/jsjkx.241100093

Abstract

PDF(2825KB) ( 8 )

References | Related Articles | Metrics

With the advancement of 5G technology,urban transportation system can facilitate information exchange more efficiently,significantly enhancing the road experience for all participants.In this system,roadside units access and forward data from on-board units,pedestrian terminals,and road sensors through 5G connectivity,providing drivers and pedestrians with effective situational awareness and reducing the incidence of accidents,thereby improving driving efficiency.However,existing authentication schemes suffer from low security,high overhead,and poor scalability.To address these issues,this paper proposes an end-to-end lightweight secure authentication and key update scheme for Internet of Vehicles.Firstly,it implements fine-grained access control for RSU permissions through factorial trees.Then,it designs a mutual authentication protocol for RSUs and IOV devices using low-cost algorithms such as physical unclonable function.Finally,it provides a dynamic online key update method for parti-cipating nodes to address issues like IoV mobility.The scheme’s security against various attacks is verified through Proverif,and informal analysis.Comparisons with other studies in terms of performance overhead demonstrate that the proposed scheme is suitable for resource-constrained IoV scenario.

Automatic Attack Path Discovery Method for Substation Remote Monitoring Network

SHI Junnan, CHEN Zemao, ZHANG Liqiang

Computer Science. 2025, 52 (12): 339-350. doi:10.11896/jsjkx.250100143

Abstract

PDF(3855KB) ( 4 )

References | Related Articles | Metrics

As substations evolve from isolated systems to complex networks spanning IT and OT,the security threats they faced are increasing,making the identification of potential attack paths in remote monitoring networks crucial.This paper presents an automated attack path planning method based on the MITRE ATT & CK framework.It treats ATT & CK techniques as attack primitives and maps attack stages using the Cyber Kill Chain.A formalized threat model is constructed,and a PDDL-based me-thod for automatic generation is proposed,transforming the attack path discovery problem into a general automated planning issue for fine-grained analysis.Experimental results show that this method reduces reliance on user expertise,generates comprehensive and practically valuable attack paths based on specific network topology,and provides strong support for automated penetration testing and security defense system development.

Rank-sorting Hybrid Genetic Algorithm for Search High Quality Balanced Boolean Functions

ZHAO Haixia, LI Xin, WEI Yongzhuang

Computer Science. 2025, 52 (12): 351-357. doi:10.11896/jsjkx.241200039

Abstract

PDF(1668KB) ( 7 )

References | Related Articles | Metrics

The balanced Boolean functions with favorable security indicators are always used as core component in symmetric cipher,which can guarantee the overall security of cipher.One of important approaches to get high quality Boolean functions is searching by using heuristic algorithm.This paper designs Rank-sorting hybrid genetic algorithm to search balanced Boolean functions with high nonlinearity and low absolute autocorrelation values.Compared to traditional genetic algorithms,the following strategies and methods are designed and used in Rank-sorting hybrid genetic algorithm.Firstly,the crossover protection strategy is designed and used in the crossover phase,which can assure the balance of offspring.Secondly,elite selection strategy based on the value of fitness function is utilized in the selection step,in order to prevent the loss of excellent individuals.In particular,a sorting method named Rank-sorting algorithm is proposed and implemented on the selected offspring before they entering the next iteration,the result of using Rank-sorting algorithm is that the differences between individuals for the next crossover are increased,the possibility of generating excellent offspring is enhanced and the risk of the whole algorithm getting stuck in local optimal solutions is reduced.Experimental results show that for the Boolean functions with even number(6 to 14) of variables,ba-lanced Boolean functions with almost optimal nonlinearity and low autocorrelation can be searched by using Rank-sorting hybrid genetic algorithm.

Research on Malicious Domain Detection Based on Heterogeneous Graph Inductive Learning

LIANG Jianpeng, MO Xiuliang, WANG Pengxiang, WANG Huanran, WANG Chundong

Computer Science. 2025, 52 (12): 358-366. doi:10.11896/jsjkx.241000083

Abstract

PDF(2688KB) ( 5 )

References | Related Articles | Metrics

Current malicious domain detection techniques based on graph neural networks rely on domain experts for meta-path selection to convert heterogeneous graphs into homogeneous graphs for direct learning.This approach struggles to leverage the rich topological information within the graph and lacks good scalability and generalization capabilities.For this issue,this paper proposes a malicious domain detection technique based on inductive learning from heterogeneous graphs.Firstly,it constructs a heterogeneous information network with nodes representing domains,hosts,and domain registration information using a meta-path generation algorithm.Secondly,to address the model’s poor applicability in real networks under direct training,it utilizes the inductive graph neural network HeteroGAT to learn the general structure of the heterogeneous graph formed by training samples and enhances detection performance through an autoencoder-based domain feature representation.Finally,it compares the proposed algorithm with machine learning and deep learning methods on public datasets.Experimental results demonstrate that the proposed method achieves superior performance metrics and effectively handles data imbalance even with a limited number of training samples,showing strong robustness.

Smart Contract Bytecode Vulnerability Detection Method Based on Heterogeneous Graphs and Instruction Sequences

SONG Jianhua, CAO Kai, ZHANG Yan

Computer Science. 2025, 52 (12): 367-373. doi:10.11896/jsjkx.241100076

Abstract

PDF(1608KB) ( 7 )

References | Related Articles | Metrics

In recent years,the security issues of smart contracts have become increasingly prominent,and vulnerability detection has become a key challenge.In scenarios where source code is not publicly available,bytecode-based detection methods have attracted significant attention.However,existing deep learning methods typically rely solely on sequences or graph structures,which makes it difficult to fully capture vulnerability features.To address this,this paper proposes a smart contract bytecode vulnerability detection method based on heterogeneous graphs and instruction sequences,called RGCN-ResNet1D(Relational Graph Convolutional Network and ResNet-based 1D Convolutional Network).This method models bytecode as a heterogeneous graph and instruction sequence,using a Relational Graph Convolutional Network(RGCN) to extract structural features and a ResNet-based 1D Convolutional Network(ResNet1D) to extract sequential features,and then fuses the two types of features for vulnerability detection.A cross-entropy loss function is also designed,which dynamically adjusts the weight based on the number of misclassified samples,effectively alleviating the class imbalance problem in the training set.Experimental results show that RGCN-ResNet1D achieves F1 scores of 95.43%,90.67%,and 92.31% for detecting integer overflow,timestamp dependency,and self-destruct vulnerabilities,respectively,significantly outperforming the comparison methods.

Highly Robust Model Structure Backdoor Method Based on Feature Distribution

CHEN Xianyi, ZHANG Chengjuan, QIAN Jiangfeng, GUO Qianbin, CUI Qi, FU Zhangjie

Computer Science. 2025, 52 (12): 374-383. doi:10.11896/jsjkx.250300064

Abstract

PDF(5058KB) ( 6 )

References | Related Articles | Metrics

Model backdoor attacks traditionally hide triggers within model parameters,activating predetermined outputs when specific samples are presented.However,such methods are vulnerable to defense techniques like parameter pruning,making backdoors difficult to trigger.This paper introduces a novel approach based on feature distribution for backdoor triggering,creating a structure-based backdoor independent of model parameters,achieving high concealment and robustness.Firstly,distribution-based triggers in the model’s feature space are used to generate backdoor images,enabling more stable backdoor activation and improving attack reliability.Secondly,a backdoor structure consisting of a distribution detector and backdoor register is embedded within target layers.This structured backdoor doesn’t rely on model parameters,significantly enhancing robustness and resis-tance to detection.Finally,the distribution detector extracts distribution-based trigger patterns while the backdoor register activates and contaminates model features,ensuring precise backdoor triggering under expected conditions for more targeted effects.Experimental results demonstrate that the proposed method maintains a 100% attack success rate even after 20 rounds of para-meter modifications and can evade multiple advanced backdoor detection mechanisms.

Ring Signature Scheme Based on Domestic Cryptographic Algorithm SM9

XIE Zhenjie, ZHANG Yao, YANG Qichao, SONG Enzhou

Computer Science. 2025, 52 (12): 384-390. doi:10.11896/jsjkx.241000072

Abstract

PDF(1582KB) ( 8 )

References | Related Articles | Metrics

Ring signatures possess spontaneity and anonymity,which can protect user privacy while implementing digital signatures.They have been widely used in scenarios requiring the concealment of the signer’s true identity,such as blockchain,electronic voting,and digital currency transactions.Identity-based cryptography can avoid the complexity of public key infrastructure construction and public key certificate management,which offers better system performance.Based on the SM9 digital signature algorithm,an identity-based domestic cryptographic algorithm,this paper constructs a ring signature scheme that satisfies general system model and security model.In the random oracle model,it is proven that this scheme has EUF-CMIA security based on the q-SDH hard problem.It is also proven that this scheme maintains full anonymity even in the presence of adversaries with access to the system’s master secret key.Theoretical analysis and testing indicates that this scheme has significant performance advantages over existing similar schemes.When the number of ring users is 1 024,the signature and verification rates are improved by 121% and 111% respectively,and the signature data size is reduced by nearly 50%.

Active Learning-based Multi-modal Fusion Rumor Detection

SHANG Yunxian, CAI Guoyong, LIU Qinghua, JIANG Yiming

Computer Science. 2025, 52 (12): 391-399. doi:10.11896/jsjkx.241000161

Abstract

PDF(2953KB) ( 5 )

References | Related Articles | Metrics

Traditional rumor detection methods still have many shortcomings,such as insufficient utilization of multi-modal information,failure to consider sample scarcity,high labeling costs,and domain shifts.Therefore,it cannot meet the demands.To address the issues of sample scarcity and domain changes,this paper proposes a new Active Learning-based Multi-modal Fusion Rumor detection model,called ALMF.ALMF designs a novel uncertainty query strategy enhanced by the propagation structure graph,which ensures that the samples selected through active learning have greater learning value and reduces the demand for sample labeling.Meanwhile,ALMF employs multi-modal data,fully integrating text features,image features,and propagation structure features.The complementary enhancement between different modal features improves the capability of rumor detection.ALMF is tested on the PHEME and WEIBO datasets.The results show that ALMF outperforms the compared models,achieving an accuracy improvement of 2% to 9%.Compared to active learning based on basic query strategies,ALMF achieves performance that is nearly equivalent to that of full sample utilization with only approximately 5% of the samples labeled.By employing a query strategy enhanced with propagation structure graphs and cross-modal enhancement fusion methods,the ALMF model successfully addresses the challenges associated with rumor detection in new domain events.

Image Encryption Algorithm Based on Novel Chaotic System and Binary Block CompressedSensing

FAN Haiju, YUE Shuang, DOU Yuqiang, LI Ming, ZHANG Mingzhu

Computer Science. 2025, 52 (12): 400-410. doi:10.11896/jsjkx.241100045

Abstract

PDF(6652KB) ( 7 )

References | Related Articles | Metrics

An image encryption scheme based on novel integer chaotic system and block compressed sensing(BCS) is proposed for embedded devices.Firstly,an integer chaotic system is designed by perturbing one-dimensional variable parameters.Then,in the compressed sensing process,the integer chaotic system is used to generate a binary measurement matrix to reduce the computation amount of embedded devices.In order to further improve the efficiency of compressed sensing,the original image is processed in blocks.Finally,the image is diffused to achieve secondary encryption.In the decryption part,PL algorithm and wiener filter are used to accelerate the decryption and improve the decryption quality of the image.Simulation experiments and performance analysis show that the proposed integer chaotic system has good chaotic characteristics.When the computer word length is 32,its Lyapunov index is always above 20.2.The image encryption scheme based on the chaotic system has high decryption quality,high security performance,and can resist common attack means.

Firmware Recovery Based Emulation and Testing Method for Industrial Gateway

WEI Zihan, MA Rongkuan, LI Beibei, YANG Yahui, LI Zhuo, SONG Yunkai

Computer Science. 2025, 52 (12): 411-418. doi:10.11896/jsjkx.241200143

Abstract

PDF(2457KB) ( 4 )

References | Related Articles | Metrics

With the continuous development of intelligent manufacturing industry,edge computing devices represented by industrial gateways are widely used in industrial sites.At the same time,software vulnerabilities of industrial gateways are beginning to affect the security of industrial networks.However,due to the specialized implementation of industrial gateway and the low-fidelity of firmware extraction,existing methods could not meet the security testing requirements.To address these issues,a firmware recovery based emulation and testing method for industrial gateway is proposed.Firstly,based on the extraction of the firmware filesystem,a heuristic recovery method is employed to free up and repair duplicate and erroneous system files,which provides a file access basis for emulation.Secondly,a heuristic emulation intervention method is adopted to mitigate errors occurring during emulation,which implements test-orientated emulation.Finally,a fuzzer is designed for industrial gateways that can be emulated.In evaluation part,firmware filesystem recovery is performed on four real industrial gateways.The emulations and fuzzing tests are conducted on important applications in two industrial gateways.The evaluation results reveal an average reduction of 27.2% in the degree of chaos for the recovered filesystem,and show a good result for emulation.Moreover,an undisclosed denial of ser-vice vulnerability in real industrial gateway devices is discovered during the fuzzing tests,which proves the effectiveness of the work.

Detection of Web Command Injection Vulnerabilities on IOS-XE Based on Static Analysis-drivenApproach

LU Bo, LYU Xiao

Computer Science. 2025, 52 (12): 419-427. doi:10.11896/jsjkx.250100060

Abstract

PDF(2364KB) ( 7 )

References | Related Articles | Metrics

Vulnerability mining for the Web interface of network devices has become very common,and the abuse of vulnerabilities poses a serious threat,the security and stability of network devices catch the attention in the security field.Fuzzing is the main method for Web interface vulnerability mining of network devices,but these methods have little effect on the Cisco IOS-XE system.Therefore,a static analysis-driven fuzzing framework based on the IOS-XE webUI,called IOXFuzzer,is proposed to detect the underlying command injection vulnerabilities.IOXFuzzer increases the probability of discovering vulnerable code by mo-delling back-end Lua scripts with abstract syntax trees,constructing dangerous path libraries to trace dangerous paths backwards,constructing parameter trees to filter high-quality seed libraries,and generating high-coverage test cases.At the end,IOXFuzzer is evaluated on Cisco ASR 1000,ISR 4000 series physical devices,and CSR 1000v series devices with 69 different firmware versions from 2019 to present and detects a total of eight underlying command injection vulnerabilities,one of which is undisclosed.

Enhancing NLP Robustness Against Attacks with Retrieval-augmented Classification and Decoupled Representations

ZHANG Peng, ZHANG Daojuan, CHEN Kai, ZHAO Yufei, ZHANG Yingjie, FEI Kexiong

Computer Science. 2025, 52 (12): 428-434. doi:10.11896/jsjkx.250500005

Abstract

PDF(2002KB) ( 10 )

References | Related Articles | Metrics

While NLP models have achieved state-of-the-art performance across various classification tasks,their vulnerability to adversarial attacks remains a significant challenge.This paper introduces a novel retrieval-augmented classification approach designed to enhance model robustness against such attacks.By leveraging KNN retrieval mechanism,this method interpolates the predicted label distributions with those of retrieved instances,strengthening the model’s decision-making process in adversarial settings.A key innovation of this work is the decoupling of the representation spaces used for classification and retrieval,which mitigates performance degradation and training instability caused by shared representations.The proposed method is evaluated across a range of benchmark datasets under various adversarial attack scenarios,demonstrating substantial improvements in model robustness.Specifically,the accuracy drops typically observed under adversarial conditions are reduced by 30 percentage points to 40 percentage points,with the proposed approach maintaining performance stability even under intense attacks.Comprehensive experiments validate the effectiveness of the proposed method,highlighting the impact of both retrieval-augmented classification and decoupled representations in creating more resilient and reliable systems.