Computer Science

Large Language Models and Rumors:A Survey on Generation and Detection

PAN Jie, WANG Juan, WANG Nan

Computer Science. 2025, 52 (11): 1-12. doi:10.11896/jsjkx.250700034

Abstract

PDF(2068KB) ( 152 )

References | Related Articles | Metrics

Rumor detection has been an interdisciplinary research topic since the mid-20th century.The rapid rise of social-media platforms such as Weibo and Twitter has kept the task in the spotlight,and the surge of rumors during the 2016 U.S.presidential election brought it to wider public attention.Breakthroughs in LLMs have dramatically advanced natural-language understanding and generation,catalyzing profound changes in the field of rumor detection.This paper presents a systematic survey of the latest studies on rumor generation and detection in the LLM era.It firstly revisits the concept of social-media rumors and summarizes widely used benchmark datasets,tracing the evolution of detection frameworks from traditional machine learning to deep learning and graph neural networks.It then analyzes in depth the four core roles that LLMs play in rumor detection,parameter fine-tu-ning,zero/few-shot prompting,knowledge augmentation and multimodal fusion.In addition,it catalogs datasets containing LLM-generated rumors and examines emerging detection techniques for AI-generated content,such as watermarking,linguistic fingerprints,and semantic-entropy-based methods.This paper concludes by outlining future research directions and the key challenges that remain.

Research on Domain Knowledge Question Answering via Large Language Models withCompositional Context Prompting

FANG Quan, ZHANG Jinlong, WANG Bingqian, HU Jun

Computer Science. 2025, 52 (11): 13-21. doi:10.11896/jsjkx.241200198

Abstract

PDF(2465KB) ( 87 )

References | Related Articles | Metrics

In recent years,the rapid development of large language models has garnered widespread attention across various sectors.While these models naturally excel at various natural language processing tasks,their performance in domain-specific question answering tasks often falls short due to a lack of specialized training in vertical domains,leading to unreliable and less applicable answers.To improve the performance of domain knowledge question answering systems,this paper proposes a novel approach based on compositional context prompting for large language models.Compositional context prompting consists of domain knowledge context and question-answer example context.The domain knowledge context is retrieved from the domain knowledge base using a contrastive learning based dense retriever,which can enhance the domain expertise processing ability of large language models.The question-answer example context is obtained through semantic similarity retrieval from the training set,which improves the large language model's understanding of question intent.Finally,the obtained composite context prompts are inputted into the large-scale language model fine-tuned with domain knowledge to generate the final domain answers.Through extensive experiments and comprehensive comparisons with baseline models,the proposed method achieves an improvement of 15.91% in precision and 16.14% in recall on the BERTScore metric compared to ChatGPT,with an F1 Score improvement of 15.87%.

Zero-shot Knowledge Extraction Method Based on Large Language Model Enhanced

PI Qiankun, LU Jicang, ZHU Taojie, PENG Yueling

Computer Science. 2025, 52 (11): 22-29. doi:10.11896/jsjkx.241000049

Abstract

PDF(4977KB) ( 89 )

References | Related Articles | Metrics

The knowledge extraction task aims to extract structured knowledge from complex information resources.However,existing research on knowledge extraction often relies on a large amount of manually annotated data,leading to high costs.To address this challenge,this paper proposes a zero-shot knowledge extraction method enhanced by large language models,which aims to perform knowledge extraction tasks automatically without relying on any manually annotated data,leveraging the strong semantic reasoning capabilities of large models to reduce data annotation costs.Specifically,it first preprocesses the format of the test set data and fine-tunes a general-purpose large model across domains to obtain a data annotation model.This model is then used to annotate relevant texts to extract corresponding entity and attribute inference information.Next,this paper establishes a new chain of thought prompting paradigm for this information and further fine-tunes a specialized large model for a specific domain to obtain a knowledge extraction model.Additionally,it continuously increases the data and iteratively trains to enhance the model's performance.Finally,it enhances the attribute information of the test set using the large model to improve the knowledge extraction model's understanding of the text,thereby enhancing its extraction performance.Benchmarking experiments on multiple large models further demonstrate that the proposed zero-shot knowledge extraction framework achieves a significant perfor-mance improvement.

DF-RAG:A Retrieval-augmented Generation Method Based on Query Rewriting and Knowledge Selection

ZHANG Haoran, HAO Wenning, JIN Dawei, CHENG Kai, ZHAI Ying

Computer Science. 2025, 52 (11): 30-39. doi:10.11896/jsjkx.241000117

Abstract

PDF(4014KB) ( 81 )

References | Related Articles | Metrics

Large language models have demonstrated formidable comprehension abilities in conversational tasks,yet they still face issues such as data timeliness and inefficiency in handling specific knowledge.To address these challenges,Retrieval-augmented Generation(RAG) has emerged as an effective solution.However,existing RAG systems still encounter significant challenges,including query understanding bias,inflexible external knowledge retrieval strategies,and low relevance of retrieval results.In response to these issues,this paper proposes a Dynamic Fine-grained Retrieval-augmented Generation(DF-RAG) approach.This method comprises three modules:a query understander,a knowledge selector,and a response generator.By rewriting queries and incorporating externally relevant documents into response generation,it enhances the retrieval-augmented large language model pipeline,achieving dynamic fine-grained retrieval augmentation.Comparative experiments and analyses are conducted on four open-domain question answering datasets against four different types of benchmarks.The results indicate that DF-RAG can more effectively integrate external knowledge with the model's inherent knowledge when handling complex and ambiguous queries.This study holds significant importance for improving the model's text retrieval and response generation capabilities in complex tasks.

Instruct-Malware:Control Flow Graph Based Large Language Model Analysis of Malware

ZHOU Yuchen, LI Peng, HAN Keji

Computer Science. 2025, 52 (11): 40-48. doi:10.11896/jsjkx.241100118

Abstract

PDF(3210KB) ( 27 )

References | Related Articles | Metrics

Malware detection and classification face challenges due to their complexity and stealthiness.Although GNNs can effectively model control flow graphs,thereby enhancing the accuracy of behavioral pattern recognition,their “black-box” nature limits interpretability.Moreover,existing methods rely heavily on large amounts of labeled data,resulting in weaker generalization capabilities and difficulties in addressing novel malware.LLMs possess strong feature extraction and contextual understanding abilities,capable of efficiently processing few-shot data and achieving multimodal information fusion,thus enhancing analytical precision and generalizability.Inspired by large language models and leveraging contrastive learning strategies,this paper aims to simultaneously learn the structure of control flow graphs and assembly instructions,thereby enhancing the effectiveness and flexibility of malware analysis.Based on this,this paper designs the Instruct-Malware framework,which employs lightweight graph-text alignment projection and two-stage instruction optimization,significantly enhancing the flexibility and robustness of malware analysis.Additionally,the interpretability of the model has been improved,clarifying the decision-makingprocess.Experimental results demonstrate that the proposed framework exhibits significant performance improvements in malware classification and subgraph recognition tasks,surpassing current mainstream approaches and substantially narrowing the gap with specialized mo-dels.This provides new insights into building an efficient and reliable malware analysis system.

Optimal Scale Combinations and Attribute Reduction for Partially Incomplete Generalized Multi- scale Data

ZHOU Shilin, WU Weizhi, LI Tongjun

Computer Science. 2025, 52 (11): 49-61. doi:10.11896/jsjkx.250700019

Abstract

PDF(1655KB) ( 22 )

References | Related Articles | Metrics

For the issue of knowledge acquisition in partially incomplete generalized multi-scale data sets,firstly,this paper pro-poses a method to transform a partially incomplete generalized multi-scale decision system into a generalized multi-scale set-valued decision one.Then,a tolerance relation on the object sets under each scale combination with each attribute subset in the obtained generalized multi-scale set-valued decision system is then constructed.Corresponding tolerance classes are also built.Upper and lower approximations,belief and plausibility degrees of sets with respect to each tolerance relation as well as information quantities of the attribute subset are subsequently presented.Six types of optimal scale combinations in a consistent generalized multi-scale set-valued decision system are further defined and their relationships are examined.It is rigorously demonstrated that five types of optimal scale combinations are equivalent while there is no static relationship between the concept of information quantity optimal scale combination with any of the other five types.Finally,an attribute reduction approach based on a belief optimal scale combination in a consistent generalized multi-scale set-valued decision system is developed,and an illustrative example is employed to explain the calculation of a belief optimal scale reduct.

Truster:Efficient Query-oriented Clustered Storage Solution for Autonomous Vehicle TrajectoryData

WANG Zhengquan, PENG Zhiyong

Computer Science. 2025, 52 (11): 62-70. doi:10.11896/jsjkx.241100052

Abstract

PDF(3078KB) ( 15 )

References | Related Articles | Metrics

Autonomous vehicle trajectory data holds significant research and practical value,attracting extensive attention to its storage and querying technologies.However,existing trajectory data management solutions are primarily designed for general trajectory data and exhibit notable limitations in efficiently writing autonomous driving trajectory data with high sampling frequency.Additionally,the high cost of index maintenance in dynamic environments makes it challenging to meet the demands of dynamic updates and real-time queries.To address the challenges of achieving high-frequency writes,dynamic updates,and real-time queries for high-sampling-rate and high-real-time trajectory data in autonomous driving scenarios,this paper proposes Truster,an efficient query-oriented clustered storage solution for autonomous vehicle trajectory data.This method includes the design of an encoder and embedder to generate position-aware keys for raw trajectories and extract feature vectors;a storage structure based on a Log-Structured Merge tree(LSM-tree)called the CLSM-tree to achieve clustered storage of similar trajectories;an LCC compaction strategy that leverages locality-sensitive hashing(LSH)for rapid clustering during the compaction of Sorted-String Tables(SSTables);and a trajectory query algorithm that uses multi-granularity cache and bucket mapping to quickly narrow down the search space.Truster not only supports high-frequency writes but also maintains index adaptability to dynamic workloads while offering enhanced query efficiency.Comparative experiments on the real-world autonomous vehicle trajectory dataset Argoverse demonstrate that Truster achieves a 20% to 200% improvement in write performance and a 20% to 100% improvement in query performance compared to existing methods.

Spatial Pyramid Bag of Words Algorithm Based on Persistent Homology

YI Lisha, PENG Ningning

Computer Science. 2025, 52 (11): 71-81. doi:10.11896/jsjkx.240900160

Abstract

PDF(8384KB) ( 28 )

References | Related Articles | Metrics

To address the mismatch between the output form of topological features extracted from persistent homology and the common input form of machine learning algorithms,this paper proposes a new algorithmic framework－Spatial Pyramid Bag of Words Algorithm Based on Persistent Homology(PHSBoW).This algorithm transforms the persistent diagrams(PDs) generated by persistent homology into fixed-length vectors while maximizing the retention of the topological features contained within the PD diagrams.To improve accuracy and reduce runtime,this paper further develops three algorithms－PHSsBoW,PHSwBoW,and PHSVLAD—based on the PHSBoW algorithm through enhancements like weight optimization,substitution with clustering mo-dels,and expansion of the bag of words model.By conducting experiments on nine datasets of varying types and scales,it combines these four algorithms with support vector machines for classification.The experimental results indicate that,compared to traditional kernel function algorithms(SWK,PSSK,PWGK) and vectorization algorithms(PBoW,PI,PL),classification accuracy is improved on average by 3.29 percentage points to 17.98 percentage points,and runtime is significantly reduced relative to kernel function algorithms.This demonstrates that these algorithms effectively address the challenges of integrating persistent homology into machine learning while significantly enhancing classification accuracy and algorithm execution speed.

Self-attention-based Graph Contrastive Learning for Recommendation

HU Jintao, XIAN Guangming

Computer Science. 2025, 52 (11): 82-89. doi:10.11896/jsjkx.240900134

Abstract

PDF(2627KB) ( 37 )

References | Related Articles | Metrics

With the explosive growth of Internet data,recommender systems have become crucial for addressing the problem of information overload.Graph contrastive learning-based recommendation models have demonstrated significant advantages in enhancing model performance by improving user-item interaction graphs.Although these models have achieved some success,most existing methods rely on perturbing graph structures for data augmentation.However,this approach struggles to preserve the inherent semantic structure and is vulnerable to noise interference.To further enhance the performance of recommendation models,this paper proposes a novel self-attention-based graph contrastive learning recommendation algorithm(AttGCL).On the one hand,the integrated self-attention mechanism strengthens the connections between users and items,allowing the model to capture user preferences and individual differences more accurately.On the other hand,the ICL loss function effectively controls the importance of positive and negative samples,leading to better alignment between global and local representations.This method retains the essential semantics of user-item interactions,enabling the model to reflect user preferences more accurately and improve recommendation effectiveness.Experimental results on three real-world datasets show that AttGCL significantly outperforms existing graph contrastive learning methods in terms of performance,demonstrating its advantages in efficiency and robustness.

Adversarial Generative Multi-sensitive Attribute Data Biasing Method

WANG Wenpeng, GE Hongwei, LI Ting

Computer Science. 2025, 52 (11): 90-97. doi:10.11896/jsjkx.240900061

Abstract

PDF(1953KB) ( 30 )

References | Related Articles | Metrics

This paper proposes a method for multi-sensitive attribute data debiasing,leveraging adversarial learning and autoencoder to eliminate correlations between sensitive and non-sensitive attributes,minimize the impact on model accuracy when striving for fairness,and address the issue of multi-sensitive attribute debiasing.In addressing multi-sensitive attribute debiasing,this method groups based on the combined values of multiple sensitive attributes,enhancing the fairness of each group's predictions by eliminating group correlations with these sensitive attribute combinations.To eliminate correlations between sensitive and non-sensitive attributes,an adversarial training approach is employed,utilizing auto-encoders alongside networks predicting sensitive attributes.This training effectively uncovers and eliminates latent sensitive attribute-related information within the groups,signi-ficantly reducing bias while retaining data utility.To mitigate the impact on model accuracy from striving for fairness and optimize the balance between accuracy and fairness,a prediction network is introduced.Its loss function is used as a constraint to enhance the encoder's ability to extract information,ensuring more precise capture of key information during data encoding and preventing excessive sacrifice of predictive performance during the debiasing process.Data debiasing experiments on three real datasets are conducted,applying the encoded data to logistic regression models.The fairness improvements range from 50.5% to 84%,validating the effectiveness of the debiasing method.Considering fairness,accuracy,and their balance,this debiasing method outperforms other debiasing algorithms.

Survey on Image Deblurring Algorithms

CHEN Kang, LIN Jianhan, LIU Yuanjie

Computer Science. 2025, 52 (11): 98-112. doi:10.11896/jsjkx.241200045

Abstract

PDF(2965KB) ( 25 )

References | Related Articles | Metrics

Image deblurring is a classic problem in computer vision,aiming to recover sharp visual information from blurry input images or videos.Blur is often caused by factors such as camera misfocus,camera shake,or fast-moving objects.Traditional deblurring methods typically model the task as a deconvolution problem,treating the blurry image as the convolution of a sharp image and a blur kernel.However,these methods face limitations when dealing with complex or non-ideal blur types.In recent years,deep learning-based regression methods have made significant breakthroughs.These approaches leverage architectures such as Convolutional Neural Networks(CNNs) and Transformers to learn the mapping between blurry and sharp images,enabling effective handling of complex blur scenarios without explicit modeling of the blur kernel.Additionally,generative deep learning methods,such as Generative Adversarial Networks(GANs) and Diffusion models,have shown considerable potential in the deblurring field.Generative AI,by modeling and learning the image detail generation process,not only effectively removes blur but also generates high-quality images with fine textures,demonstrating superior performance in challenging blur scenarios.This paper first introduces the characteristics of image blur and outlines common deblurring tasks and evaluation metrics.It then delves into the fundamental architectures and training methods of deblurring models,providing a comparative analysis of representative state-of-the-art deblurring models.Finally,the paper explores potential future research directions in the field of image deblurring.

FE-DARFormer:Image Desnowing Model Based on Frequency Enhancement and Degradation- aware Routing Transformer

QIN Yi, ZHAN Pengxiang, XIAN Feng, LIU Chenlong, WANG Minghui

Computer Science. 2025, 52 (11): 113-122. doi:10.11896/jsjkx.241200176

Abstract

PDF(4218KB) ( 23 )

References | Related Articles | Metrics

The goal of image desnowing is to restore clear scene information from images degraded by complex snowy scenes.Unlike the regularity and semi-transparency of rain,snow exhibits various forms and scales of degradation,with severely degraded regions often obstructing important scene details.Recent methods have employed self-attention mechanisms to address different degradation phenomena.However,global self-attention computation across all image regions is computationally expensive,leading these methods to restrict attention to smaller windows.Yet,due to the occlusion effects in severely degraded areas,the recovery of these regions relies heavily on capturing information from surrounding areas,which results in a receptive field bottleneck,limi-ting the ability to aggregate sufficient information.As a result,these methods struggle to effectively restore large-scale degraded regions.To improve desnowing performance,this paper proposes a novel approach,introducing a new network architecture called FE-DARFormer,which combines a Degradation-Aware Routing Transformer and a Dual-Frequency Enhancement Transformer.FE-DARFormer dynamically routes and applies global self-attention to severely degraded regions,enabling a global receptive field for effective restoration of large degraded areas while reducing computational cost.Additionally,it uses discrete wavelet decomposition to handle multi-scale snow degradation,enhancing the recovery of diverse snowflake shapes and textures.

Infrared and Visible Image Fusion Cross-modality Contrastive Representation Network Based on Rolling MLP Feature Extraction

YAN Zhilin, NIE Rencan

Computer Science. 2025, 52 (11): 123-130. doi:10.11896/jsjkx.240800110

Abstract

PDF(4111KB) ( 20 )

References | Related Articles | Metrics

At present,in the fusion of infrared and visible image,dataset lacks the real fusion image to guide the important diffe-rence information of the two modes required for the final fusion image.Most of the existing fusion methods only consider the tradeoff and interaction of the source image,ignoring the role of the fusion image in the fusion process.The important information in the fusion image can constrain the difference information of the source image.Therefore,this paper proposesCRN to better guide the extraction of important information in the source image required by the fused image.At the same time,improving the quality of fusion image reconstruction can further strengthen the guidance of important feature information of source image.The quality of the reconstructed image is related to the extracted features.Among existing feature extraction methods,CNN has poor performance in capturing global features,while Transformer has high computational complexity and poor learning ability of local features.On this basis,a CNN module D2 Block combined with MLP is introduced,which can effectively extract and fuse local features and remote dependencies by rolling feature mappings in different directions.A large number of qualitative and quantitative experiments on several public data sets show that the proposed method achieves better results than other advanced methods.

UAV Small Object Detection Algorithm Based on Feature Enhancement and Context Fusion

CHEN Chongyang, PENG Li, YANG Jielong

Computer Science. 2025, 52 (11): 131-140. doi:10.11896/jsjkx.241000017

Abstract

PDF(3576KB) ( 21 )

References | Related Articles | Metrics

Aiming at the problems of low detection accuracy caused by small object sizes,insufficient feature information,dense distribution,and occlusion in UAV aerial photography,this paper proposes a UAV small object detection algorithm based on feature enhancement and context fusion.Firstly,a lightweight backbone network for enhanced feature extraction is constructed,utilizing lightweight feature extraction blocks to efficiently extract feature information,and a fine-grained channel fusion block is designed to effectively prevent the loss of fine-grained features.The backbone network improves the feature extraction capability and inference speed of the model.Secondly,a small object detection head is constructed to fully extract the position information and detailed features of small objects.Then,the adaptive spatial attention module is used to adaptively adjust the receptive fields required for different objects,making full use of the rich context information around the aerial small objects.Finally,a minimum point distance-based bounding box regression loss function(MPDIoU) is introduced to further improve the precision of dense small object detection.The proposed algorithm achieves mAP_0.5 and mAP_0.5:0.95 of 46.7% and 28.6% on the VisDrone2019 dataset,respectively,representing an improvement of 8.5% and 5.9% over the baseline network YOLOv8s.Moreover,the algorithm reduces parameters by 23.4% compared to YOLOv8s,making it efficient for dense small object detection in UAV aerial photography scenarios.

Human-Object Interaction Detection Based on Fine-grained Attention Mechanism

DING Yuanbo, BAI Lin, LI Taoshen

Computer Science. 2025, 52 (11): 141-149. doi:10.11896/jsjkx.240900113

Abstract

PDF(3046KB) ( 18 )

References | Related Articles | Metrics

Fine-grained information,as a kind of contextual information,can assist models in recognizing human-object interactions with similar relative spatial relationships.However,how to utilize this key cue to uniformly model feature information of different granularities on multi-scale feature maps remains a critical challenge that hinder further improvement of human-object interaction detection accuracy.To address this problem,this paper proposes a human-object interaction detection model based on fine-grained attention mechanism.The model strengthens local features under the guidance of fine-grained information.It fuses feature maps of different scales and automatically learns image content through a deformable attention mechanism.Additionally,it models the long-range dependencies between features of various granularities,essentially improving the accuracy of the human-object interaction detection model.Extensive experiments are conducted on the V-COCO and HICO datasets.The experimental results show that the proposed method has increased the mAPby 7.7 percentage points on the V-COCO dataset,and the mAP has increased by 7.43,7.5 and 7.85 percentage points on the HICO dataset compared to the baseline models.

Video Compressed Sensing Method with Integrated Deformable 3D Convolution and Transformer

DU Xiuli, ZHU Jinyao, GAO Xing, LYU Yana, QIU Shaoming

Computer Science. 2025, 52 (11): 150-156. doi:10.11896/jsjkx.240800026

Abstract

PDF(3102KB) ( 25 )

References | Related Articles | Metrics

Facing the challenge of increasing data volume due to higher resolution of video,realizing high quality video reconstruction with lower sampling rate can reduce the consumption of communication resources and thus reduce the difficulty of deployment at the sampling end.However,the existing video compressed sensing methods cannot fully utilize the inter-frame correlation of the video,and the reconstruction quality of the video at low sampling rates needs to be further improved.With the introduction of deep learning technology,distributed video compression sensing based on deep learning provides new ideas for video compression sensing reconstruction.Therefore,this paper combines 3D deformable convolution with Transformer to construct CS3Dformer network,which utilizes the effectiveness of 3D deformable convolutional network in capturing local and spatio-temporal features of video and learns spatio-temporal features between video frames,and at the same time,utilizes the advantages of Transformer in capturing long-range dependency features,which compensates to some extent for the advantages of convolutional neural network method in capturing the non-local similarity of the defects of image,and better realize the modeling of the video.This method is an end-to-end video compression perception method,the experimental results on multiple datasets verify the effectiveness of the proposed method.

Multi-level Feature Fusion Image Emotion Recognition Based on Color Enhancement

LI Xiaoyu, QIAN Yi, WEN Yimin, MIU Yuqing

Computer Science. 2025, 52 (11): 157-165. doi:10.11896/jsjkx.241000016

Abstract

PDF(4600KB) ( 24 )

References | Related Articles | Metrics

Image emotion recognition aims to analyze and understand the emotion conveyed by the content of images.The primary challenge lies in bridging the gap between latent visual features and abstract emotion.Existing deep learning methods mostly bridge this gap by extracting different levels of features,but they often overlook the importance of color features.To address the problem,this paper proposes a multi-level feature fusion image emotion recognition method based on color enhancement.By introducing a color enhancement module and a multi-level feature extraction module,more representative feature representations are extracted.The color enhancement module extracts color features from both the RGB and HSV color spaces using the color moment method,and expands their dimensions to enrich emotional information.The multi-level feature extraction module introduces an attention mechanism to focus on key regions in the image and employs decision fusion to mitigate the issue of information redundancy caused by concatenating high-level and low-level features.Experiments conducted on four public datasets demonstrate that the proposed method can effectively recognize image emotion and significantly improve performance compared to mainstream image emotion recognition methods.

Gesture Action Prediction Based on VMD Composite Neural Network Model

ZHAO Lian, WU Yangdong, DENG Zhifang, LI Fengshuo, YUAN Qingni, ZHANG Taihua

Computer Science. 2025, 52 (11): 166-174. doi:10.11896/jsjkx.241000115

Abstract

PDF(3549KB) ( 21 )

References | Related Articles | Metrics

sEMG is often used to predict human intention behavior.It is an unstable,non-periodic and noisy bioelectrical signal,which is easily affected by power frequency interference and environmental interference,so it is difficult to predict it.This paper proposes CNNM based onVMD and PSO algorithm.The model combines LSTM,CNN andBiLSTM.Firstly,the Ninapro dataset is used to optimize the parameters of VMD through the improved PSO algorithm,and the sEMG signal is processed by VMD.The decomposed components are weighted and reconstructed according to the Hilbert energy method,which reduces the complexity of the signal and retains the key features.Then,the LSTM method is used to extract the temporal features from sEMG signals,the CNN method is used to further extract the spatial features,and the attention mechanism is used to strengthen the extraction of key information.Finally,it is input into BiLSTM for prediction and recognition.Experimental results show that the prediction accuracy of the proposed model can reach 99.9%,and the prediction accuracy of CNNM is improved by 3%～8% compared with other models.Finally,the role of each module is verified by ablation experiments.This research aims to improve the prediction and recognition accuracy of gesture actions and provide an effective solution for the control of rehabilitation training robots.

Neural Radiance Field for Human Reconstruction Based on Multi-scale Hierarchical Network

WANG Yang, WANG Guodong, ZHAO Junli, SHENG Xiaomeng

Computer Science. 2025, 52 (11): 175-183. doi:10.11896/jsjkx.240900141

Abstract

PDF(3156KB) ( 16 )

References | Related Articles | Metrics

The reconstruction of 3D human models from monocular RGB video faces challenges in accurately capturing human poses,especially when using prior models like SMPL.Due to its rigid assumptions,such models struggle to depict subtle pose variations,leading to suboptimal reconstruction results.Additionally,existing NeRF-based human modeling methods often generate unnatural shadows or floating artifacts around certain body parts when rendering unseen poses,and their representation of texture details tends to be insufficient.To address these issues,this paper proposes a hierarchical network based on the Triplane Multiscale learning,aims at enhancing the texture details of 3D human models through NeRF techniques and improving the model's generalization capability across different poses.In terms of methodology,multi-resolution hash encoding is employed to replace the traditional sinusoidal frequency encoding function,allowing for more efficient capture of high-frequency human features and speeding up model convergence.The Triplane Multiscale learning strategy is applied to capture pose details,effectively improving the accuracy and visual quality of 3D reconstructions.Experiments demonstrate that the proposed improvements significantly enhance the reconstruction of 3D human models,especially when handling complex pose variations.The method shows notable advantages in terms of training speed,rendering quality,and pose generalization capabilities.By applying this model,the resulting 3D human models exhibit more realistic details,and the synthesized results for novel poses are of high quality,further advancing the development of 3D human reconstruction technology from monocular video.

Object Detection Based on Deep Feature Enhancement and Path Aggregation Optimization

WANG Xiaofeng, HUANG Junjun, TAN Wenya, SHEN Zixuan

Computer Science. 2025, 52 (11): 184-195. doi:10.11896/jsjkx.241100107

Abstract

PDF(5434KB) ( 19 )

References | Related Articles | Metrics

In deep networks,the feature information of the input data is gradually abstracted and compressed during the feed-forward process,resulting in some of the feature information that is crucial for object detection being diluted or lost.Based on YOLOv11n,an object detection method with deep feature enhancement and path aggregation optimization is proposed. Firstly,GLFEM is designed to combine the local features of the feature map with the global features to strengthen the expression ability of the deep network features.Then,AFEM is designed to dynamically enhance the feature extraction ability of the deep network according to the reliability of the features. Finally,the path aggregation feature pyramid network is optimized to fuse the feature information between different levels and reduce the semantic information difference between levels.Experimental results on three public datasets,VisDrone,NWPU VHR-10,and TinyPerson,show that the average detection accuracy of the proposed method is improved compared to current state-of-the-art object detectors.Experiments on the self-built dataset AirportTiny also show the proposed method achieves good performance,it has good generalisation ability.

Medical Image Target Detection Method Based on Multi-branch Attention and Deep Down- sampling

GU Chengjie, MENG Yi, ZHU Dongjun, ZHANG Junjun

Computer Science. 2025, 52 (11): 196-205. doi:10.11896/jsjkx.240900088

Abstract

PDF(3563KB) ( 20 )

References | Related Articles | Metrics

With the development of artificial intelligence technology,medical image detection based on deep learning has a wide application prospect in clinical practice.However,for some medical image target detection such as tumor and plaque,there are some problems,such as small area to be labeled,few features to be extracted and difficult to extract.To solve these problems,this paper proposes a medical image target detection method(MD-Det) based on multi-branch attention and deep subsampling.The feature extraction module(C2f-DWR) is introduced to extract multi-scale features and enhance the feature representation of the target.This paper designes a deep down-sampling module(D-down) to capture the context information in the image more effectively and enhance the feature extraction capability.The core idea is to combine average pooling and maximum pooling operations to make full use of their respective advantages to improve the feature extraction effect by fusing multiple sampling methods.The accuracy of target detection is improved while maintaining the computational efficiency.Then,a multi-branch attention(MA) mechanism is proposed,which extracts and weights features of different dimensions,with each branch extracting features of different dimensions of the input tensor,including spatial and channel features.By generating attention weights,important features are emphasized and weighted together.The feature extraction capability of the network is enhanced,and the detection perfor-mance of the model is improved.Finally,a new joint optimization strategy is proposed,which weights Wise-IoU loss and NWD loss to form a joint regression loss function to further improve the accuracy of target recognition.Experiments show that the proposed method can effectively improve the detection accuracy of the model in medical image targets,and the mAP_0.5of the medical data sets Tumor and Liver are increased by 2.5 percentage points and 1.1 percentage points,respectively.

Image Deraining Based on Union Attention Mechanism and Multi-stage Feature Extraction

LIN Zukai, HOU Guojia, WANG Guodong, PAN Zhenkuan

Computer Science. 2025, 52 (11): 206-212. doi:10.11896/jsjkx.240900013

Abstract

PDF(4405KB) ( 18 )

References | Related Articles | Metrics

Existing image deraining networks predominantly rely on the large-scale synthetic paired datasets for training,ignoring the difference in spatial distribution characteristics and the difference in channel importance between synthetic and real data,resulting in blurred texture details and diminished generalization performance.To address these issues,this paper develops an unsupervised network model for image deraining based on a union attention mechanism with multi-stage feature extraction.To adapt to the spatial locality of rain streaks,the feature-aware module is initially designed to extract rain streaks through the combination of spatial and channel attention mechanisms,while dilation convolution is used to enhance the sensory field of rain feature extraction.In addition,a recurrent neural network is introduced to extract the rain stripe features gradually,and the useful information of the previous stage is retained in the cycle to improve the rain stripe feature extraction ability.To further enhance the discrimination of local micro-details and global texture structure features,it designs a multi-scale discriminator for distinguishing images at three different scales and guideings the generator to produce higher quality images.Qualitative and quantitative experiments on synthetic and real datasets show that the proposed method is superior to some supervised,semi-supervised and unsupervised me-thods on PSNR,SSIM and NIQE metrics,which verifies its effectiveness and generalization.

Multi-source Domain Generalization Fault Diagnosis Method Based on Instance-level PromptGeneration

LI Shugang, LI Mingjia, YUAN Longhui, QI Guangpeng, LIU Chi

Computer Science. 2025, 52 (11): 213-222. doi:10.11896/jsjkx.250300117

Abstract

PDF(2961KB) ( 18 )

References | Related Articles | Metrics

This paper proposes multi-source domain generalization fault diagnosis method based on instance-level prompt generation to enhance the model's fault recognition capability in cross-domain environments.This method employs a cross-frequency aligned prompt generator to dynamically generate instance-level prompts,enabling refined modeling of local features across diffe-rent samples.It incorporates a semantic consistency enhancement module to ensure the semantic validity of instance-level prompts.Furthermore,to improve the model's stability and adaptability in cross-domain tasks,a memory bank-enhanced contrastive learning module is introduced,which fully utilizes cross-domain positive and negative samples.By storing and dynamically updating feature representations of training samples,this module expands the diversity of positive and negative sample distributions and enhances the effectiveness of cross-domain feature learning.Additionally,a FourierMix module is adopted to perform frequency-domain feature mixing of samples from different source domains,dynamically generating simulated samples to strengthen the model's adaptability on unseen target domains.Experimental results on CWRU and Paderborn datasets demonstrate that the proposed method outperforms existing approaches across multiple unseen target domains,achieving average classification accuracies of 93.54%(1.52% improvement over state-of-the-art) on CWRU dataset and 90.52% (1.30% improvement) on Paderborn dataset.Experimental results prove its effectiveness and robustness in industrial fault diagnosis tasks.

Method for Generating Judgment Documents Based on Trial Logic

LIAO Jinchao, YANG Weizhe, QIN Yongbin, HUANG Ruizhang, CHEN Yanping, ZHOU Yulin

Computer Science. 2025, 52 (11): 223-229. doi:10.11896/jsjkx.250500054

Abstract

PDF(2520KB) ( 18 )

References | Related Articles | Metrics

The automatic generation of judicial documents is one of the key tasks in the construction of smart courts,aiming to enhance judicial efficiency and document quality.However,due to the blind spots of large models in judicial cognition,they struggle to understand the trial mechanism and document norms,resulting in deficiencies in the logical consistency and structural rationality of the generated documents.To address these issues,this paper proposes a method for generating judicial documents based on trial logic,which utilizes large language models to simulate the trial reasoning process and generate documents in stages.Firstly,legal elements are used to fill in the preset template to describe the “basic case facts”.Secondly,the facts and evidence are analyzed and aligned to obtain the “trial facts”.Finally,relevant legal provisions are retrieved from the knowledge base to gene-rate the “court judgment”,and the complete document is assembled.Experimental results show that,compared with the baseline model on real case file data,the proposed method has improved the F1 values of ROUGE-1,ROUGE-2,and ROUGE-L by 6.03,6.56,and 7.98 percentage points respectively,verifying the effectiveness of the proposed method.

Relationship and Attribute Aware Entity Alignment Based on Variant-attention

LI Zhikang, DENG Yichen, YU Dunhui, XIAO Kui

Computer Science. 2025, 52 (11): 230-236. doi:10.11896/jsjkx.240800140

Abstract

PDF(2532KB) ( 23 )

References | Related Articles | Metrics

When distinguishing the different representation effects of different neighbors on the central entity,existing entity alignment methods mostly use feature similarity or local feature information of the relationships between entities to calculate attention coefficients.Those methods ignore the global information of relationships,and cannot effectively distinguish the effect of different information on entity alignment.To address this problem,this paper proposes an entity alignment model,which assigns different weights to different types of relationships based on the frequency of(relationship,neighbor) appearing in the entire graph,then integrates the weights into GAT to obtain a variant attention mechanism for aggregating different neighbors.Meanwhile,different attribute values are aggregated in a similar way by using the frequency information of(attributes,attribute values) in the entire graph.After that,combining the structure and entity name embedding with two types of information respectively to obtain two embedding representations of the central entity.Finally,entity alignment is performed based on the weighted distance between the two embedding representations,aims to distinguish the different effects of relationship and attribute information on the entity alignment.Experimental results show that this model is superior to other methods in the three cross lingual datasets of DBP15K,Hit@1 increases by a maximum of 2.15% compare to the optimal method,besides,results also show significant improvement in both Hit@10 and MRR,which indicate that the proposed model can effectively enhance the accuracy of entity alignment.

Commonsense Question Answering Model Based on Graph-Text Integrating

CAI Ruixiang, ZHAO Shuliang, HE Jiayao

Computer Science. 2025, 52 (11): 237-244. doi:10.11896/jsjkx.240900081

Abstract

PDF(2234KB) ( 26 )

References | Related Articles | Metrics

Knowledge graphs have demonstrated significant effectiveness in commonsense question answering.Existing methods typically utilize entities from the question to retrieve local subgraphs from the knowledge graph(KG),which are then encoded using graph neural networks(GNN).Subsequently,the GNN-encoded results are combined with language models(LMs) to infer answers and answer the questions.However,commonsense question answering systems using GNNs and LMs face two challenges:1) how to efficiently extract subgraphs from the knowledge graph,effectively represent and utilize their knowledge and structural information; 2) how to achieve deep integration and joint reasoning of the question context and subgraph knowledge.This paper proposes a graph-text integrating model for commonsense question answering(Graph-Text Integrating for Commonsense Question Answering,GTICQA).The model initially refines key entities by filtering through an external dictionary,achieving pruning of the knowledge subgraph,and then separately encodes the question context using an LM and the refined knowledge subgraph using a GNN encoder.Additionally,during the subgraph encoding process,a novel k-sparse attention mechanism is introduced to enhance the extraction of global features from the subgraph and suppress noise.Finally,a knowledge fusion method that includes fine-grained bimodal interaction fusion layers and mean interaction fusion layers is used to deeply integrate and dynamically update the two knowledge representations.The GTICQA model is evaluated on three datasets:CommonsenseQA,OpenBookQA,and MedQA-USMLE,achieving accuracy rates of 79.12%,72.20%,and 39.40%,respectively,surpassing the current best methods,demonstrating the model's advantage in handling commonsense question answering.

Multi-strategy Improved Electric Eel Foraging Optimization Algorithm

WANG Xinwei, FENG Feng

Computer Science. 2025, 52 (11): 245-254. doi:10.11896/jsjkx.241100106

Abstract

PDF(5746KB) ( 25 )

References | Related Articles | Metrics

In response to the issues of EEFO algorithm,such as insufficient global exploration ability,susceptibility to local optima,slow convergence,and performance sensitivity to parameter settings that require careful adjustment and optimization,a multi-strategy improved Electric Eel Foraging Optimization algorithm(IEEFO)is proposed.Firstly,the energy factor strategy is adjusted by introducing a hyperbolic tangent energy factor,which allows the algorithm to incorporate exploratory behavior earlier in the iteration process,enabling rapid discovery of the optimal population and accelerating convergence speed.Secondly,thedistur-bance factor is improved to increase the range of positions where the electric eel can move,which is beneficial for global optimization of the population.Then,a sine cosine strategy is added during the migration phase,which is conducive to local exploration of the algorithm.Finally,after each iteration,a lens imaging reverse learning strategy is incorporated to expand the search space,which helps the algorithm escape from local optima and accelerate convergence to the global optimal solution.The IEEFO is compared with 6 basic algorithms and 4 single-strategy improved Electric Eel Foraging Optimization algorithms,and 13 benchmark functions are used for simulation experiments to evaluate the performance of the IEEFO algorithm.The experimental results show that the IEEFO has faster convergence speed and stronger global optimization ability compared to the aforementioned algorithms,with a significant improvement in overall algorithm performance.Additionally,a mechanical optimization design experiment is conducted to further test and analyze the effectiveness and applicability of the IEEFO.

Review of Blockchain Consensus Algorithm

ZHOU Kai, CHEN Fu, LU Tianyuan, CAO Huaihu

Computer Science. 2025, 52 (11): 255-269. doi:10.11896/jsjkx.241100140

Abstract

PDF(3970KB) ( 20 )

References | Related Articles | Metrics

The consensus algorithm is a critical technological cornerstone of blockchain,facilitating consistency among nodes within a distributed system concerning specific data.The primary bottleneck in current consensus algorithms lies in the impact of communication complexity on blockchain performance,specifically in terms of latency and throughput.To address these challenges,this paper presents a comprehensive analysis the development of consensus technology,with a particular focus on the Basic-Round(BR)-based Directed Acyclic Graph(DAG) classification criterion.The present study aims to analyse the core principles of the BR-DAG consensus algorithm and the consensus process.The objective of this analysis is to mitigate the inherent limitations of BR-DAG consensus algorithms by reducing network communication latency,enhancing convergence speed,and increasing transaction throughput.This study offers an extensive review of the current state of research,existing challenges,and emerging trends in advanced consensus algorithms,with a specific emphasis on BBCA-Chain.Furthermore,we propose a robust evaluation framework designed to facilitate the comparative analysis of various consensus algorithms based on throughput,latency,and other performance dimensions,aligned with established classification criteria.Finally,it discusses the prevailing challenges faced by consensus algorithms and proposes future research directions that should focus on BR-DAG and Rho-calculate.This includes the development of concurrent computation models based on message interaction delivery and the formal verification of consensus algorithm correctness.To achieve a high-throughput,low-latency,and robust consensus algorithm,formal verification methods can be employed.

SCDDA:SCA and Dinkelbach-based Approach for UAV Trajectory and Computation Offloading in Space-Air-Ground Integrated Networks

ZHENG Jingjing, CHEN Xing, ZHANG Jianshan

Computer Science. 2025, 52 (11): 270-279. doi:10.11896/jsjkx.241100163

Abstract

PDF(3174KB) ( 26 )

References | Related Articles | Metrics

The massive amount of heterogeneous data generated by the widespread use of mobile devices has placed higher demands on data communication networks.In this context,the sixth-generation(6G) mobile network is expected to meet the needs of various mobile devices executing computation-intensive and latency-sensitive mobile applications.Currently,the novel Space-Air-Ground Integrated Network(SAGIN),which results from the organic combination of network components in space,air,and ground,has become a key component of the 6G architecture.Compared with traditional terrestrial communication paradigms,SAGIN can effectively enhance the coverage and throughput of mobile communication networks by utilizing non-terrestrial network components such as satellites,high-altitude platforms,and UAVs.This makes it well-suited to meet the needs of a large number of mobile devices in infrastructure-less areas.Despite the significant potential of SAGIN in various aspects for infrastructure-less regions,its application still faces practical challenges such as resource constraints,dynamic changes in network topology,and service requirements of mobile devices.To address these challenges,this paper considers the impact of user mobility on system energy efficiency under real-world scenarios,and investigates a joint optimization problem of computation offloading and UAV trajectory in SAGIN.To solve the targeted joint optimization problem,an efficient and novel algorithm based on convex optimization techniques is designed,decoupling the target problem into two sub-problems.These sub-problems are solved separately using the SCA and the Dinkelbach method,to obtain an approximate optimal solution to the target optimization problem.Numerical simulation results demonstrate that the proposed algorithm outperforms other benchmark algorithms in terms of performance.

Multipath Routing Algorithm for Satellite Networks Based on Convolutional Twin Delay Deep Deterministic Policy Gradient

WEI Debin, ZHANG Yi, XU Pingduo, WANG Xinrui

Computer Science. 2025, 52 (11): 280-288. doi:10.11896/jsjkx.240800161

Abstract

PDF(3703KB) ( 17 )

References | Related Articles | Metrics

In the satellite network,due to the influence of geographical location and people's living habits,the difference in the needs of users in the satellite coverage area will cause the load imbalance of the satellite network.A multi-path routing algorithm based onconvolutional double-delay deep deterministic policy gradient(CTD3-MR) is proposed for the above problem.Under the SDN structure,CTD3 is deployed in the controller as the agent,and the dynamically changing links' residual bandwidth,transmission delay,packet loss rate and spatiotemporal level are trained as the network state input agent,and the output action is used as the network link weight,and the weighted sum of the maximum link bandwidth utilization,average end-to-end delay and network packet loss rate is used as the reward function to adjust the action.After the agent training converges,the controller obtains the k-shortest path according to the network link weight output by the agent,and takes the path weight ratio as the path traffic allocation ratio to generate an optimal routing strategy and forward it to the satellite for multipath transmission.Finally,CTD3-MR is compared with TD3,TMR and ECMP routing algorithms.Experimental results show that compared with other routing algorithms,CTD3-MR reduces the average end-to-end delay by at least 7.64%,the packet loss rate by 28.65%,the maximum link bandwidth utilization by 11.44%,and the traffic distribution index by 5.82%,which improves the network load balancing performance.

Virtual Coordinate-based Deployment Strategy for Low-orbit Satellite Caching

ZHANG Tai, DU Shu, CHEN Shaolei

Computer Science. 2025, 52 (11): 289-297. doi:10.11896/jsjkx.240900001

Abstract

PDF(3945KB) ( 19 )

References | Related Articles | Metrics

In the swiftly advancing domain of satellite communication networks,a significant challenge arises concerning the efficient distribution of massive content volumes and optimal resource utilization.To tackle this issue,an in-network caching solution grounded in ICN has been proposed,in order to improve the immediacy and efficiency of content retrieval.Despite these advancements,low-earth orbit(LEO)satellites face constraints due to their limited onboard storage capabilities and energy consumption,making traditional caching strategies suboptimal.This paper introduces an innovative deployment methodology for cache nodes within LEO satellites using virtual coordinates.The approach begins by assessing real-world demand,segmenting service areas according to the density of content requests from terrestrial users.This segmentation enables the precise identification of caching demand hotspots.To simplify the intricate spatial relationships involved,a hierarchical mapping technique is utilized,translating these complexities into a more manageable mathematical framework.Building upon this foundation,a weighted clustering algorithm is then applied to refine the placement of cache nodes,thereby optimizing the allocation of caching resources across the sa-tellite network.Empirical evaluations affirm that this method not only significantly decreases user content acquisition latency but also achieves an energy consumption reduction exceeding 15%.These enhancements underscore the potential of this strategy to elevate the operational performance of the system.In critical scenarios,such as power grid emergencies,the pre-positioning of essential emergency data facilitates swift command dissemination and repair coordination amid communication disruptions,bolstering the reliability of power supply maintenance.Consequently,this approach not only aids in expediting societal stabilization efforts but also heralds extensive application prospects.For example,in the event of a power grid failure,pre-stored emergency instructions can expedite recovery operations,thereby mitigating broader societal impacts.Additionally,the enhanced content accessibility coupled with lower energy expenditure positions this strategy as a pivotal advancement in the realm of satellite communication networks.

Joint Optimization of Pilot Design and Channel Estimation Based on Deep Learning

WANG Anyi, LI Ruoman, LI Xinyu, LI Mingzhu

Computer Science. 2025, 52 (11): 298-305. doi:10.11896/jsjkx.241000004

Abstract

PDF(4060KB) ( 17 )

References | Related Articles | Metrics

With the continuous innovation and development of mobile communication technology,higher requirements have emerged communication reliability and data transmission performance.Accurate and efficient acquisition of CSI is a key prerequisite for fully harnessing the technical potential of wireless communication systems.In response to the challenges of high pilot overhead and low channel estimation accuracy in MIMO-OFDM systems,a joint optimization scheme of pilot design and channel estimation based on Deep Learning(AE-DRSN) is designed.The scheme initially utilizes Concrete AE to identify and select the pilot positions with maximum information content to achieve pilot optimization.The optimized pilot positions are then input into the DRSN to achieve more accurate CSI and further complete the accurate estimation of the channel.Experimental results demonstrate that compared with traditional channel estimation methods,the joint optimization scheme based on AE-DRSN can still achieve high-precision channel estimation with minimal pilot overhead,fully verifying the effectiveness of the scheme.

Performance Optimization of Wireless Edge Storage System Based on SDN and Drone Assistance in Disaster Scenarios

SUN Shiquan, YE Miao, ZHU Cheng, WANG Yong, JIANG Qiuxiang

Computer Science. 2025, 52 (11): 306-319. doi:10.11896/jsjkx.240900004

Abstract

PDF(3483KB) ( 25 )

References | Related Articles | Metrics

Traditional edge distributed storage systems often suffer from cumbersome network configuration and high operational overhead in measuring network state information.During peak demand periods for data storage and retrieval by terminal devices,network links can become overloaded,adversely affecting data transmission performance.Furthermore,existing distributed sto-rage systems typically consider only the remaining storage space of nodes when selecting storage nodes,neglecting the impact of network state and node load on system storage performance.To address these issues,this paper designs and implements an edge-distributed storage system assisted by software-defined network(SDN) and drones.The system uses SDN technology to measure network state,node load,and storage node load information.Drones fly above heavily loaded network nodes to offload traffic and balance the load across different links.For the selection of heavily loaded network nodes and storage nodes,this paper proposes a node selection algorithm based on a multi-attribute decision model that comprehensively considers network state and node load.The algorithm identifies heavily loaded network nodes and suitable storage nodes,and deployment of drones helps achieve traffic offloading and load balancing.Experimental tests on a wireless Mesh network topology demonstrate that the proposed wireless edge-distributed storage system outperforms existing edge-distributed storage systems in terms of storage performance.The proposed system significantly reduces storage time and maintains good performance even under increased traffic load,demonstrating excellent load-balancing capabilities.

MDGRec:Multi-relation Aware Third-party Library Recommendation with Dual Graph NeuralNetworks for Mobile Application Development

CHEN Yuhan, WANG Jian, LI Duantengchuan, ZHENG Chao, LI Bing

Computer Science. 2025, 52 (11): 320-329. doi:10.11896/jsjkx.241200129

Abstract

PDF(3313KB) ( 19 )

References | Related Articles | Metrics

Third-party library(TPL)recommendation systems are designed to help developers select suitable libraries,thus improving the efficiency of mobile application(App) development.Most existing methods based on graph neural networks typically propagate information for both App and TPL nodes within a single heterogeneous interaction graph,leading to issues of data imbalance and feature confusion,limiting the recommendation accuracy.Moreover,these methods often fail to account for the complex relationships inherent in the context of TPL recommendation.To overcome these limitations,this paper proposes a multi-relation aware third-party library recommendation method with dual graph neural networks for mobile application development(MDGRec).The model employs a dual graph structure to separately model Apps and TPLs,generating distinct embeddings.Based on this,the model incorporates multiple relationships and uses adaptive weights to capture the contribution of each relation in information propagation,constructing fine-grained node representations.Experimental results on two real-world datasets show that the proposed model surpasses mainstream baselines across all metrics.

Adaptive Android Program Test Method Based on Thompson Sampling

ZHAO Yingnan, LENG Chongyang, HAN Qilong, YU Cheng

Computer Science. 2025, 52 (11): 330-338. doi:10.11896/jsjkx.240900150

Abstract

PDF(1995KB) ( 19 )

References | Related Articles | Metrics

Recently,the research of Android graphical interface test method has attracted wide attention.At present,most test methods are developed based on reinforcement learning.However,the existing methods can explore the application by selecting parameters according to experience,and can not to change the parameter settings according to the interface changes adaptively.This paper proposes an adaptive Android testing method based on Thompson sampling,which combines Thompson sampling with the Q-learning algorithm,enabling the agent to adaptively determine the next exploration action based on the current state of the interface controls being examined so that to balance exploitation and exploration more effectively for superior testing outcomes.Firstly,events of the jump at the interface during exploration are modeled for the Beta probability distribution,then a probability distribution matrix is obtained,which is weighted-averaged with the Q-table.It can take into account both the exploration value and the exploition value of the events.At the same time,the probability distribution of operational events under the current interface is sampled,and the maximum sampling value is the exploration probability value,combined with the weighted matrix,it can guide the test more comprehensively,so as to realize the adaptive exploration of the Android application interface.Experiments executed on 13 Android applications confirm the efficacy of the proposed method through experimentalcomparison and analysis with conventional reinforcement learning testing tools.

Research on Optimization of Test Case Generation Based on Neuron Coverage Index

XIAO Ziqin, SHI Yaqing, QU Yubin

Computer Science. 2025, 52 (11): 339-348. doi:10.11896/jsjkx.240900006

Abstract

PDF(2416KB) ( 13 )

References | Related Articles | Metrics

DNNs have been widely applied in many fields,and testing them is particularly important due to their complexity and uncertainty.Traditional testing methods rely too much on a single indicator and cannot fully reveal the complete behavioral patterns of deep neural networks.Therefore,it is necessary to comprehensively consider different coverage indicators to more comprehensively evaluate the performance of the model.It combines six multi-granularity deep neural network coverage metrics,optimizes the mutation strategy and seed selection steps of fuzzy testing,generates high-quality and high-coverage test cases.Experi-ments are conducted on four models of different complexities on the MNIST and CIFAR10 datasets.The original training set and newly generated effective test cases are combined for retraining the model to classification accuracy.The experimental results show that this method can significantly improve coverage and classification accuracy by optimizing the model through adaptive retraining.

Survey of Adversarial Attack and Defense for RBG and Infrared Multimodal Object Detection

ZHENG Haibin, LIN Xiuhao, CHEN Jingwen, CHEN Jinyin

Computer Science. 2025, 52 (11): 349-363. doi:10.11896/jsjkx.241200151

Abstract

PDF(2950KB) ( 15 )

References | Related Articles | Metrics

Object detection,as a fundamental classic task in the field of computer vision,has a wide range of applications.Deep learning based object detection algorithms have become the mainstream of current research due to their superior performance.However,most object detection algorithms only perform single-mode detection on visible or infrared images.In general,visible images have poor imaging in harsh weather,nighttime,and scenes,where targets are obstructed,leading to a decrease in detection performance.The use of infrared images can improve the above issues,but infrared images may miss some details of the target.Therefore,multimodal fusion detection algorithms based on visible light and infrared images are gradually emerging.However,existing research has focused on improving the performance of multimodal object detection algorithms,and research on their security is relatively scattered.Based on existing research work,this paper provides an overview of the security of multimodal object detection in adversarial situations.Firstly,a theoretical analysis of multimodal object detection and attack and defense is conducted.Secondly,multimodal object detection methods are classified and summarized according to fusion detection in different time periods.Then,existing methods of object detection and adversarial defense are summarized and organized,and the existing dataset and main evaluation indicators of multimodal object detection are summarized.Finally,potential research directions for multi-modal object detection in the future are discussed,further promoting the development and application of multimodal object detection in adversarial security research.

Research Status and Challenges of Eavesdropping Attacks and Defenses Targeting VoiceAssistants

HUANG Wenbin, REN Ju, CAO Hangcheng, JIANG Hongbo, XIONG Lizhi, CHEN Xianyi, FU Zhangjie

Computer Science. 2025, 52 (11): 364-372. doi:10.11896/jsjkx.250300047

Abstract

PDF(2441KB) ( 29 )

References | Related Articles | Metrics

Voice assistants serve as convenient interfaces for human-computer voice interaction,finding widespread application across various settings including homes,sports,and vehicles.They play a pivotal role in facilitating the intelligent advancement of industries,such as healthcare,finance,and education.However,the widespread adoption and convenience of voice assistants have also precipitated significant concerns regarding the eavesdropping on user conversations,consequently leading issues of user privacy disclosure.Existing literature primarily focuses on voice spoofing attacks and defenses,as well as adversarial sample attacks and defenses.However,there remains a notable gap in the analysis and synthesis of voice eavesdropping attacks and defenses.To address this gap,this work delves deeply into the mechanisms of eavesdropping attacks on voice assistants and meticulously reviews existing research in this domain.Firstly,this work conducts a comprehensive review and in-depth analysis of various eavesdropping attack methods,categorizs them based on their implementation strategies.It explores the means of attack,targets,necessary technology and permissions,and concealment techniques employed,aiming to provide a comprehensive understanding of the potential threats faced by voice assistants.Secondly,recent research efforts aimed at defending against voice assistant eavesdropping attacks are systematically reviewed.Through the classification and summarization of different defense technologies,coupled with insights into their application scenarios and detection effectiveness,the paper highlights the shortcomings and challenges of existing defense mechanisms,thereby offering valuable insights for enhancing the security of voice assistants.Lastly,this study meticulously analyzes the primary research challenges in the realm of eavesdropping attacks and defenses,while also discussing potential future research directions.By identifying these challenges and proposing future avenues of exploration,the paper aims to guide ongoing research endeavors towards bolstering the resilience of voice assistant systems against eavesdropping threats.

Intelligent Botnet Traffic Detection Method Based on Multi-granularity Statistical Features

ZHANG Haixia, HUANG Kezhen, LIAN Yifeng, ZHAO Changzhi, YUAN Yunjing, PENG Yuanyuan

Computer Science. 2025, 52 (11): 373-381. doi:10.11896/jsjkx.241100019

Abstract

PDF(3070KB) ( 23 )

References | Related Articles | Metrics

With the rapid development of information technology,botnet attacks have become a higly harmful cyber security threat.Botnet detection and disposal can prevent attackers from launching other derivative attacks based on botnets.The current botnet detection methods have limitations such as single feature selection perspective,easy to be bypassed or high false alarm rate.In response to these limitations,this paper proposes an intelligent botnet traffic detection method based on multi-granularity statistical features.This method extracts local coarse-grained statistical features of the network flows to be detected and global fine-grained profile of the source IP based on historical network flows,and then uses the long-short term memory networks with a multi-head attention mechanism to mine the difference in these features between benign network flows and botnet flows at different times.The botnet is ultimately identified based on these differences.Comparative experiments are conducted on the CTU-13 and ISCX botnet datasets,the proposed method achieves more than 99% in accuracy,precision,recall and F1 score.

Dynamic Analysis Based Fuzz Testing for Memory Safety Vulnerabilities

YIN Jiale, CHEN Zhe

Computer Science. 2025, 52 (11): 382-389. doi:10.11896/jsjkx.241000003

Abstract

PDF(1800KB) ( 22 )

References | Related Articles | Metrics

Systems written in C often contain potential memory vulnerabilities.Fuzz testing integrated with dynamic analysis tools can uncover memory vulnerabilities but introduce significant performance overhead.Meanwhile,current popular fuzz testing me-thods focus more on improving overall code coverage,while efficiently triggering memory vulnerabilities in already covered code is also an important capability.To this end,the dynamic analysis tool Movec is improved and combined with AFL,with the innovative work primarily using pointer metadata to guide fuzz testing for efficient memory vulnerability detection.The core steps include using a source-level hash table and secondary tree to manage pointer metadata to reduce page faults caused by combining fuzz testing with dynamic analysis.Then it removes coverage instrumentation of dynamic analysis code at the assembly level to reduce the impact of redundant instrumentation on coverage calculation.Subsequently it adds minimum pointer boundary distance and memory allocation peak indicators to guide fuzz testing in efficiently detecting buffer overflows and memory allocation failure vulnerabilities,and finally optimizes the seed queue filtering logic to streamline queue size and prioritize seeds related to memory vulnerabilities.Experiments on CVE programs show that the execution throughput of Movec combined with AFL is 54% of native,while Asan and Msan are only 10% and 4%,respectively.Compared with advanced fuzzers,it can uncover memory vulnerabilities in a shorter time,reducing the time consumption by an average of 48.4%.

Using Ring Blind Signature＋Arbitration Authentication Mixed Coin Scheme

FANG Zhipeng, LI Xiaoyu

Computer Science. 2025, 52 (11): 390-397. doi:10.11896/jsjkx.241000048

Abstract

PDF(2191KB) ( 16 )

References | Related Articles | Metrics

Blockchain is a distributed ledger technology with the characteristics of decentralization,non-tampering,and data disclosure.However,data disclosure has led to the security risk of privacy leakage in the blockchain.The introduction of the mixing center as an intermediary cuts off the connection between the transferor and the receiver,and achieves the purpose of protecting the privacy of both parties to the transaction,but it still has some security loopholes,such as the mixing center can still grasp this association,the mixing center may forge the transfer,and the transferor may deny the transaction.Therefore,a quorum-blind signature based on arbitration authentication mixing technology is proposed,which uses ring-blind signatures to solve the problem of association between the mixing center,and uses arbitration authentication to solve the problem of the mixing center and user violations.Compared with the traditional coin mixing scheme,it has incomparable advantages,which can well solve the problems existing in the traditional coin mixing scheme,and has the characteristics of anonymity,non-repudiation,non-forgery,anti-Dos,etc.,which improves the traditional coin mixing service and can further protect user privacy.The response time of this scheme is positively correlated with the number of users and the number of mixing centers.Compared to the Mixcoin and Blindcoin schemes,the response time is slightly longer,but shorter than that of Coinjoin and Coinshuffle schemes.Additionally,compared to other schemes,this scheme can effectively prevent deception by users and mixing centers,better protecting transaction privacy.

Data Trusted Sharing Scheme Based on Consortium Blockchain

LIU Zhanghui, LIN Zhexu, CHEN Hanlin, MA Xinjian, CHEN Xing

Computer Science. 2025, 52 (11): 398-407. doi:10.11896/jsjkx.241000169

Abstract

PDF(3067KB) ( 32 )

References | Related Articles | Metrics

With the advent of the big data era,securing and trustworthily sharing data on an open,dynamic,and difficult-to-control Internet has become an urgent problem to solve.Blockchain can reasonably be introduced into the trust resolution mechanism for data sharing,leveraging its significant advantages in decentralization and tamper resistance.Thus,a data trusted sharing scheme based on consortium blockchain is proposed.Firstly,a consortium blockchain-based data architecture is defined to solve the problem of heterogeneous data sources and domains.Through standardized registration processes,data resources are integra-ted efficiently.Secondly,a trusted data traceability mechanism is designed and implemented to ensure the security and integrity of data flow among data requesters,compute nodes,and data providers by leaving the traces of the data sharing process on consor-tium blockchain.In addition,a data-processing-as-a-service data sharing framework is developed to support key steps in data sharing:demand matching,data sharing,and satisfaction evaluation,addressing trust challenges during the data sharing process.The experimental results show that,compared with traditional data sharing schemes,the proportion of additional time in the proposed scheme decreases to less than 30% of the total time cost as the dataset size increases.Additionally,the average latency for querying smart contracts remains stable between 0.12 and 0.2 seconds,while the average latency for writing smart contracts stays consistent at 3 to 5 seconds.

Additively Homomorphic Encryption Scheme Based on Domestic Cryptographic Algorithm SM9

XIE Zhenjie, LIU Yiming, YIN Xiaokang, LIU Shengli, ZHANG Yongguang

Computer Science. 2025, 52 (11): 408-414. doi:10.11896/jsjkx.241100188

Abstract

PDF(1464KB) ( 29 )

References | Related Articles | Metrics

In the cloud computing environment,traditional encryption schemes not only protect data confidentiality but also cause the ciphertext to lose its computability.Homomorphic encryption solves this contradiction and has been widely applied in privacy computing fields such as data aggregation,secure multi-party computing,and federated learning.Based on the encryption algorithm of the domestic cryptographic algorithm SM9,an identity-based encryption scheme with additive homomorphism property is constructed.The correctness and additive homomorphism of the scheme are carefully derived.Starting from the q-BCAA1 and DDH difficulty problems,the scheme is proven to have IND-CPA security.And the improved message recovery algorithm is described in detail.Test results show that the encryption efficiency of the proposed additively homomorphic encryption scheme increases by 42% compared to the similar scheme,and the decryption efficiency increases by 20% to 62%.

Lightweight Privacy-preserving Mobile Sensing Classification Framework Based on AddictiveSecret Sharing

HE Yuyu, ZHOU Feng, TIAN Youliang, XIONG Wei, WANG Shuai

Computer Science. 2025, 52 (11): 415-424. doi:10.11896/jsjkx.241100101

Abstract

PDF(2762KB) ( 16 )

References | Related Articles | Metrics

To address data privacy leakage in deploying convolutional neural network models on mobile sensing devices,as well as the challenge of excessive communication overhead caused by server interaction computations in privacy-preserving target classification frameworks,a lightweight privacy-preserving mobile sensing object classification framework(LPMS) based on additive secret sharing is proposed.This framework ensures that mobile sensing devices maintain data confidentiality during data exchanges while significantly reducing both communication and computational overhead.Firstly,a series of secure computing protocols are developed using additive secret sharing technology,avoiding reliance on computationally intensive cryptographic primitives to facilitate efficient and secure neural network computations.Secondly,a three-dimensional chaotic encryption scheme is introduced to protect original data from potential attackers during uploads to the edge server.Finally,the correctness and security of the LPMS framework are validated through theoretical analysis and security proofs.Experimental results demonstrate that,compared to the PPFE scheme,the LPMS framework reduces model computation overhead by 73.33% and communication overhead by 68.36%.

Neural Network Backdoor Sample Filtering Method Based on Deep Partition Aggregation

GUO Jiaming, DU Wentao, YANG Chao

Computer Science. 2025, 52 (11): 425-433. doi:10.11896/jsjkx.240900007

Abstract

PDF(3616KB) ( 18 )

References | Related Articles | Metrics

Deep neural networks are vulnerable to backdoor attacks,where attackers can implant backdoors and hijack model behavior by poisoning data.Among them,class-specific attacks can bypass most defense methods due to their complex mapping relationships and close association with normal tasks,making them more threatening.This paper studies the relationship between attack success rate and model classification performance in the process of implanting backdoors for class-specific attacks,summarizes three properties,and designs a sample filtering method based on these properties to address class-specific attacks.This methoduses the Deep Partition Aggregation(DPA) ensemble learning method and voting method to iteratively filter backdoor samples.This paper mathematically proves the effectiveness of this filtering method based on three properties of class-specific attacks,and conducts extensive experiments on standard classification datasets.After four iterations,it filters more than 95% of backdoor samples in all experiments.At the same time,the results of comparative experiments with the latest sample filtering methods,demonstrate the superiority of proposed method in addressing class-specific attacks.The experiments in this paper are based on the open-source project backdoorbox on Github.

Backdoor Attack Method for Federated Learning Based on Knowledge Distillation

ZHAO Tong, CHEN Xuebin, WANG Liu, JING Zhongrui, ZHONG Qi

Computer Science. 2025, 52 (11): 434-443. doi:10.11896/jsjkx.250100146

Abstract

PDF(4582KB) ( 23 )

References | Related Articles | Metrics

Federated learning enables different participants to jointly train a global model using their private datasets.However,the distributed nature of federated learning also provides room for backdoor attacks.The attacker of the backdoor attack poisons the global model causing the global model misleads to targeted incorrect predictions when encountering samples with specific backdoor triggers.This paper proposes a backdoor attack method for federated learning based on knowledge distillation.Firstly,the teacher model is trained using the concentrated poison dataset generated by distillation,and the “dark knowledge” of the teacher model is transferred to the student model to refine the maliciousneurons.Then,the neurons with backdoors are embedded into the global model through Z-scoreranking and mixing of neurons .The experiment is evaluated the performance of KDFLBD in iid and non-iid scenarios on common datasets.Compared with pixel attacks and label flipping attacks,KDFLBD significantly improves the attack success rate(ASR) while ensuring that the main task accuracy(MTA) is not affected.

Research on Individual Unmanned Aerial Vehicles Identification Technology Based on Voiceprint Characteristics

ZHANG Meng, QIAO Jinlan

Computer Science. 2025, 52 (11): 444-451. doi:10.11896/jsjkx.250300079

Abstract

PDF(3752KB) ( 20 )

References | Related Articles | Metrics

With the rapid development of artificial intelligence and communication technologies,unmanned aerial vehicles(UAVs) are increasingly being applied across various industries.In the field of low-altitude logistics and transportation,UAVs demonstrate significant potential due to their efficiency,convenience,and low cost.However,legitimate UAVs performing delivery tasks are highly susceptible to spoofing attacks.Shippers relying solely on visual characteristics,especially when malicious third parties use UAVs of the same model as legitimate ones,find it difficult to accurately determine whether an incoming UAV is authorized to pick up a package.To effectively address this issue,an individual UAV identification system based on voiceprint features is proposed.Firstly,audio of a hovering UAV is recorded using a mobile device,and empirical wavelet transform is applied to remove high-frequency noise from the UAV's audio signal,thereby improving the signal-to-noise ratio.Then,a filter bank is designed based on the spectral characteristics of audio signals from different UAVs of the same model,enabling efficient extraction of key audio features.Finally,a long short-term memory network model incorporating the Open-Max algorithm is constructed to handle open-set classification problems,further enhancing the system's recognition capability.Experimental results demonstrate that the proposed system achieves an accuracy of 99.8% in identifying individual UAVs of the same model and a success rate of 99.5% in detecting unauthorized UAVs,effectively mitigating spoofing attacks.