Computer Science

Review of Software Engineering Techniques and Methods Based on Explainable Artificial Intelligence

XING Ying

Computer Science. 2023, 50 (5): 3-11. doi:10.11896/jsjkx.221100159

Abstract

PDF(1672KB) ( 3819 )

References | Related Articles | Metrics

In terms of information processing and decision-making,artificial intelligence(AI) methods have shown superior performance compared to traditional methods.However,when AI models are put into production,their output results are not guaranteed to be completely accurate,so the “unreliability” of AI technology has gradually become a major obstacle to the large-scale implementation of AI.As AI is gradually applied to software engineering,the drawbacks of over-reliance on historical data and non-transparent decision-making are becoming more and more obvious,so it is crucial to provide reasonable explanations for the decision results.This paper elaborates on the basic concepts of explainable AI(XAI) and the evaluation of explanation models,and explores the feasibility of combining software engineering with explainable AI.Meanwhile,it investigates relevant researches in software engineering,analyzes the four typical application directions of XAI,namely,malware detection,high-risk component detection,software load distribution,and binary code similarity analysis,to discuss how to reveal the correctness of the system output,thereby increasing the credibility of the software system.This paper also gives insights into the research direction in combining software engineering and explainable artificial intelligence.

Study on Interpretable Click-Through Rate Prediction Based on Attention Mechanism

YANG Bin, LIANG Jing, ZHOU Jiawei, ZHAO Mengci

Computer Science. 2023, 50 (5): 12-20. doi:10.11896/jsjkx.221000032

Abstract

PDF(2857KB) ( 3612 )

References | Related Articles | Metrics

Click-Through Rate(CTR) prediction is critical to recommender systems.The improvement of CTR prediction can directly affect the earnings target of the recommender system.The performance and interpretation of the CTR prediction algorithm can guide developers to understand and evaluate recommender system accurately.That's also helpful for system design.Most existing approaches are based on linear feature interaction and deep feature extraction,which have poor model interpretation in the outcomes.Moreover,very few previous studies were conducted on the model interpretation of the CTR prediction.Therefore,in this paper,we propose a novel model which introduces multi-head self-attention mechanism to the embedding layer,the linear feature interaction component and the deep component,to study the model interpretation.We propose two models for the deep component.One is deep neural networks(DNN) enhanced by multi-head self-attention mechanism,the other computes high-order feature interaction by stacking multiple attention blocks.Furthermore,we calculate attention scores and interpret the prediction results for each component.We conduct extensive experiments using three real-world benchmark datasets.The results show that the proposed approach not only improves the effect of DeepFM effectively but also offers good model interpretation.

Explainable Comparison of Software Defect Prediction Models

LI Huilai, YANG Bin, YU Xiuli, TANG Xiaomei

Computer Science. 2023, 50 (5): 21-30. doi:10.11896/jsjkx.221000028

Abstract

PDF(2501KB) ( 3521 )

References | Related Articles | Metrics

Software defect prediction has become an important research direction in software testing.The comprehensiveness of defect prediction directly affects the efficiency of testing and program operation.However,the existing defect prediction is based on historical data,and most of them cannot give a reasonable explanation for the prediction process.This black box prediction process only shows the output results,making it difficult for people to know the impact of the internal structure of the test model on the output.In order to solve this problem,it is necessary to select software measurement methods and some typical deep lear-ning models,make a brief comparison of their input,output and structure,analyze them from the two perspectives of the degree of data differences and the processing process of the model on the code,and explain their similarities and differences.Experiments show that the method of deep learning is more effective than traditional software measurement methods in defect prediction,which is mainly caused by their different processing processes of raw data.When using convolution neural network and long-term and short-term memory neural network to predict defects,the data difference is mainly caused by the integrity of the understan-ding of code information.To sum up,in order to improve the prediction ability of software defects,the calculation of the model should comprehensively involve the semantics,logic and context of the code to avoid the omission of useful information.

Study on Reliability Prediction Model Based on BASFPA-BP

LI Honghui, CHEN Bo, LU Shuyi, ZHANG Junwen

Computer Science. 2023, 50 (5): 31-37. doi:10.11896/jsjkx.220900283

Abstract

PDF(2178KB) ( 3181 )

References | Related Articles | Metrics

Software reliability prediction is based on software reliability prediction model,which analyzes,evaluates and predicts software reliability and reliability-related measures.Using the failure data collected in software operation to predict the future software reliability.It has become an important means to evaluate software failure behavior and guarantee software reliability.BP neural network has been widely used in software reliability prediction because of its simple structure and few parameters.How-ever,the prediction accuracy of the software reliability prediction model built based on the traditional BP neural network cannot reach the expected target.Therefore,this paper proposes a software reliability prediction model based on BASFPA-BP.This model utilizes software failure data and utilizes BASFPA algorithm to optimize network weights and thresholds in the training process of BP neural network.Thus,the prediction accuracy of the model is improved.In this paper,three groups of public software failure data are selected,and the mean square error between the actual value and the predicted value is taken as the measurement standard of the predicted results.Meanwhile,BASFPA-BP is compared with FPA-BP,BP and Elman models.Experimental results show that the software reliability prediction model based on BASFPA-BP achieves high prediction accuracy in the same type of model.

Interpretable Repair Method for Event Logs Based on BERT and Weak Behavioral Profiles

LI Binghui, FANG Huan, MEI Zhenhui

Computer Science. 2023, 50 (5): 38-51. doi:10.11896/jsjkx.220900030

Abstract

PDF(4298KB) ( 3197 )

References | Related Articles | Metrics

In practical business processes,low-quality event logs due to outliers and missing values are often unavoidable.Low-quality event logs can degrade the performance of associated algorithms for process mining,which in turn interferes with the correct implementation of decisions.Under the condition that the system reference model is unknown,when performing log anomaly detection and repair work,the existing methods have the problems of needing to manually set thresholds,do not understand what behavior constraints the prediction model learns,and poor interpretability of repair results.Inspired by the fact that the pre-trained language model BERT using the masking strategy can self-supervise learning of general semantics in text through context information,combined with attention mechanism with multi-layer and multi-head,this paper proposes the model BERT4Log and weak behavioral profiles theory to perform an interpretable repair process for low-quality event logs.The proposed repair method does not need to set a threshold in advance,and only needs to perform self-supervised training once.At the same time,the method uses the weak behavioral profiles theory to quantify the degree of behavioral repair of logs.And combined with the multi-layer multi-head attention mechanism to realize the detailed interpretation process about the specific prediction results.Finally,the performance of the proposed method is evaluated on a set of public datasets,and compared with the current research with the best performance.Experimental results show that the repair performance of BERT4Log is better than the comparative research,and at the same time,the model can learn weak behavioral profiles and achieve detailed interpretation of repair results.

Review on Interpretability of Deep Learning

CHEN Chong, CHEN Jie, ZHANG Hui, CAI Lei, XUE Yaru

Computer Science. 2023, 50 (5): 52-63. doi:10.11896/jsjkx.221000044

Abstract

PDF(1759KB) ( 4004 )

References | Related Articles | Metrics

With the explosive growth of data volume and the breakthrough of deep learning theory and technology,deep learning models perform well enough in many classification and prediction tasks(image,text,voice and video data,etc.),which promotes the large-scale and industrialized application of deep learning.However,due to the high nonlinearity of the deep learning model with undefined internal logic,it is often regarded as a “black box” model which restricts further applications in key fields(such as medical treatment,finance,autonomous driving).Therefore,it is necessary to study the interpretability of deep learning.Firstly,recent studies on deep learning,the definition and necessity of explaining deep learning models are overviewed and described.Secondly,recent studies on interpretation methods of deep learning,and its classifications from the perspective of intrinsic interpretable model and attribution-based/non-attribution-based interpretation are analyzed and summarized.Then,the qualitative and quantitative performance criteria of the interpretability of deep learning are introduced.Finally,the applications of deep learning interpretability and future research directions are discussed and recommended.

Code Embedding Method Based on Neural Network

SUN Xuekai, JIANG Liehui

Computer Science. 2023, 50 (5): 64-71. doi:10.11896/jsjkx.220100094

Abstract

PDF(3927KB) ( 3227 )

References | Related Articles | Metrics

There are many application scenarios for code analysis and research,such as code plagiarism detection and software vulnerability search.With the development of artificial intelligence,neural network technology has been widely used in code analysis and research.However,the existing methods either simply treat the code as ordinary natural language processing,or use much more complex rules to sample the code.The former processing method is easy to cause the loss of key information of the code,while the latter can make the algorithm to be too complicated,and the training of the model will take a lot of time.Alon proposed an algorithm named Code2vec,which has significant advantages compared with previous code analysis methods.But the Code2vec still has some limitations.Therefore,a code embedding method based on neural network is proposed.The main idea of this method is to express the code function as the code embedding vector.First,a code function is decomposed into a series of abstract syntax tree paths,then a neural network is used to learn how to represent each path,and finally all paths are aggregated into an embedding vector to represent the current code function.A prototype system based on this method is implemented in this paper.Experimental results show that compared with Code2vec,the new algorithm has the advantages of simpler structure and faster training speed.

Hybrid Algorithm of Grey Wolf Optimizer and Arithmetic Optimization Algorithm for Class Integration Test Order Generation

ZHANG Wenning, ZHOU Qinglei, JIAO Chongyang, XU Ting

Computer Science. 2023, 50 (5): 72-81. doi:10.11896/jsjkx.220200110

Abstract

PDF(2591KB) ( 3210 )

References | Related Articles | Metrics

Integration testing is an essential and important part in software testing.Determining the orders in which classes should be tested during the object oriented integration testing is a very complex problem.The search based approaches have been proved to be efficient in generating class integration test orders(CITO),with the disadvantage of slow convergence speed and low optimization accuracy.In the grey wolf optimizer(GWO) algorithm,wolves are likely to be located in the same or certain regions,thus easily being trapped into local optima.Arithmetic optimization algorithm(AOA) is a new meta heuristic technique with excellent randomness and dispersibility.To improve the performance of CITO generation,a hybrid optimization algorithm of GWO and AOA(GWO-AOA) is proposed,combining the rapid convergence speed of GWO and strong ability to avert local optima stagnation of AOA.In the GWO-AOA,the main hunting steps of GWO is unchanged and the leading individual of AOA is replaced by the center of dominant wolfs,providing a proper balance between exploration and exploitation.In addition,random walk scheme is adopted based on the random local mutation to improve the global search ability.Experimental results indicate that the proposed method can generate promising class integration test orders with less time compared to other comparative methods.

Mechanical Equipment Fault Diagnosis Driven by Knowledge

DONG Jiaxiang, ZHAI Jiyu, MA Xin, SHEN Leixian, ZHANG Li

Computer Science. 2023, 50 (5): 82-92. doi:10.11896/jsjkx.221100160

Abstract

PDF(4275KB) ( 3326 )

References | Related Articles | Metrics

With the rapid development of social economy,modern industry now presents a trend featuring complex research objects,informationalized application methods and diversified production modes.Industrial fault diagnosis,as one of the most important research areas in modern industry,is still facing a series of technical bottlenecks due to the complexity of mechanical equipment and the lack of referential knowledge.In order to solve the above problems,this paper proposes a knowledge-driven fault diagnosis scheme for mechanical equipment,which mainly includes two parts——knowledge construction and diagnosis process.In terms of knowledge construction,this paper presents a domain knowledge graph construction method.In terms of diagnosis process,this paper designs a general mechanical equipment fault diagnosis process consisting of four steps,fault inquiry,fault location,fault cause location and fault maintenance guidance.To date,the scheme has been actually applied in a large excavator maintenance provider in China,and its effectiveness has been verified.Experimental results indicate the scheme improves the know-ledge and intelligent level of excavator fault diagnosis and shows high accuracy and practicability.The application of the scheme will be further promoted in the industry.

Review of Intelligent Device Fault Diagnosis Based on Deep Learning

HUANG Xundi, PANG Xiongwen

Computer Science. 2023, 50 (5): 93-102. doi:10.11896/jsjkx.220500197

Abstract

PDF(3901KB) ( 3629 )

References | Related Articles | Metrics

Intelligent fault diagnosis applies deep learning theory to equipment fault diagnosis,which can automatically identify the health state and fault type of equipment,and has attracted extensive attention in the field of equipment fault diagnosis.Intelligent equipment fault diagnosis realizes equipment fault diagnosis by building end-to-end AI models and algorithms to associate equipment monitoring data with machine health status.However,there are many models and algorithms for equipment fault diagnosis,but they are not common to each other.Using models that are inconsistent with monitoring data for fault diagnosis will lead to a significant decline in diagnosis accuracy.In order to solve this problem,based on the comprehensive investigation of the relevant literature of equipment fault diagnosis,this paper first briefly describes the model framework of in-depth equipment fault diagnosis,then classifies,lists,compares and summarizes the models and algorithms according to the specific application scenarios and equipment monitoring data types,and finally analyzes the future development direction according to the existing problems.This review is expected to provide a useful reference for the research of intelligent equipment fault diagnosis.

Deep Learning-based Heterogeneous Information Network Representation:A Survey

WANG Huiyan, YU Minghe, YU Ge

Computer Science. 2023, 50 (5): 103-114. doi:10.11896/jsjkx.220800112

Abstract

PDF(4570KB) ( 680 )

References | Related Articles | Metrics

Things in the nature connect mutually.There are various associations between them in the real world.For example,social networks can be constructed by the user-user relationships.The article-author relationship can be used to construct a citation network.In homogeneous networks,nodes or edges are all in the same type,resulting in a lot of information loss.In order to ensure the integrity and richness of information to a greater extent,researchers have proposed heterogeneous information network(HIN),a network model containing multiple types of nodes or edges.By embedding the topological structure and semantic information of HIN into a low-dimensional vector space,downstream tasks can utilize the rich information in the HIN for machine learning or data mining.This paperfocuses on the HIN-based representation learning tasks,and summarizes the recent representation learning methods of HIN which are based on deep learning models.We focus on two main issues:semantics extraction of HIN and information preserving of dynamic HIN.We also illustrate some new applications of HIN-based representation learning,and propose the future development trend of heterogeneous information networks.

Dataspace:A New Data Organization and Management Model

FAN Shuhuan, HOU Mengshu

Computer Science. 2023, 50 (5): 115-127. doi:10.11896/jsjkx.220700042

Abstract

PDF(3055KB) ( 1030 )

References | Related Articles | Metrics

With the rapid development of the digital economy,how to realize multi-party data fusion in an untrusted environment and find new ways for data sharing,data analysis and data services in cross-organizational scenarios has become a new problem in the upgrading of social digital industries.Dataspace brings new ideas to solve these problems.The development history of data organization and management is reviewed,and it points out that in the background of big data,systematic research on dataspace is urgent and important.The connotation of dataspace is analyzed and a formal description is given.A big data platform architecture based on dataspace is proposed,and three classic application scenarios are briefly described.Focusing on the construction of dataspace,it analyzes the current correlation research issues and main technical methods from data modeling,dynamic evolution,data query processing,security and privacy,and briefly describes the realization and application of dataspace in different fields.Finally,the research outlook and challenges are prospected from the perspective of multimodal data fusion,efficient query processing,safe data sharing,and the construction of a big data platform based on dataspace.

Deep Cross-modal Information Fusion Network for Stock Trend Prediction

CHENG Haiyang, ZHANG Jianxin, SUN Qisen, ZHANG Qiang, WEI Xiaopeng

Computer Science. 2023, 50 (5): 128-136. doi:10.11896/jsjkx.220400089

Abstract

PDF(2113KB) ( 583 )

References | Related Articles | Metrics

Stock trend prediction,as a classic and challenging task,can help traders make trading decisions for greater returns.Recently,deep learning related models have achieved obvious performance improvement on this task.However,most of the current deep learning related works only leverage the historical data on stock price to complete the trend prediction,which cannot capture the market dynamics other than price indicators,thus having an accuracy limitation to a certain extent.To this end,this paper combines social media texts with stock historical price information,and proposes a novel deep cross-modal information fusion network(DCIFNet) for stock trend prediction.DCIFNet first utilizes temporal convolution operations to encode stock prices and twitter texts,so that each element can have sufficient knowledge of its neighborhood elements.Then,the results are fed into a transformer-based cross-modal fusion structure to fuse stock prices and important information in Twitter texts more effectively.Finally,a multi-graph attention convolutional network is introduced to describe the interrelationships among different stocks,which well captures the industry,wiki and correlation relationship among related stocks,leading to the accuracy improvement of stock prediction.We have performed trend prediction and simulated trading experiments on high-frequency trading datasets in nine different industries,and ablation studies as well as compared experiments with multipronged attention network for stock forecasting(MAN-SF) demonstrate the effectiveness of the proposed DCIFNet method.In addition,with the optimal accuracy of 0.6309,it obviously outperforms representative methods on the stock prediction application.

Cost-sensitive Multigranulation Approximation of Neighborhood Rough Fuzzy Sets

YANG Jie, KUANG Juncheng, WANG Guoyin, LIU Qun

Computer Science. 2023, 50 (5): 137-145. doi:10.11896/jsjkx.220500268

Abstract

PDF(6193KB) ( 556 )

References | Related Articles | Metrics

Multigranulation neighborhood rough sets are a new data processing mode in the theory of neighborhood rough sets,in which the target concept can be characterized by upper/lower approximate boundaries of optimistic and pessimistic,respectively.Nevertheless,the current multigranulation neighborhood rough sets not only lacks the method of using the existing information granules to describe the target concept approximately,but also can not deal with the situation that the target concept is fuzzy Whereas the approximation theory of rough sets proposed by professor Zhang provides a method for approximately describing knowledge utilizing existing information granules,therefore,it provides a new method for constructing approximate and accurate sets of multigranulation neighborhood rough fuzzy sets.In this paper,aiming to process the fuzzy target concept,the approximation theory of rough sets is applied to the field of neighborhood rough sets,and a cost-sensitive approximate representation model of neighborhood rough fuzzy sets is introduced.Then,from the perspective of multigranulation,a multigranulation approximate representation model of the cost-sensitive neighborhood rough fuzzy sets is constructed with evaluating its related properties.Finally,simulation results show that when the multigranulation cost-sensitive approximation and upper/lower approximation are used to approximate the fuzzy target concept,the multigranulation cost-sensitive approximation method reaches the least misclassification cost.

Pseudo-abnormal Sample Selection for Video Anomaly Detection

ZHAO Song, FU Hao, WANG Hongxing

Computer Science. 2023, 50 (5): 146-154. doi:10.11896/jsjkx.220400227

Abstract

PDF(2926KB) ( 562 )

References | Related Articles | Metrics

Unsupervised video anomaly detection methods generally use normal video data to train an anomaly detection model through frame reconstruction or frame prediction.However,normal videos usually contain a large number of background frames and similar scenes,which are quite redundant,leading to inefficient modeling for video anomaly detection.To address this issue,this paper proposes a pseudo-abnormal sample selection method,which iteratively selects video frames with high abnormal scores from original videos to build a new concise training pool for video anomaly detection based on future frame prediction.As for the detection model,this paper designs a two-path U-Net architecture,where each path has a different sampling frequency on video frames so that spatial-temporal features of videos can be better extracted and utilized from multiple scales.In the two-path U-Net,each layer shares a memory module to strengthen the impact of typical training data for future frame prediction and video anomaly detection.Experimental evaluation on benchmark video datasets demonstrates the efficiency and effectiveness of the proposed method.

Hyperspectral Image Classification Based on Swin Transformer and 3D Residual Multilayer Fusion Network

WANG Xianwang, ZHOU Hao, ZHANG Minghui, ZHU Youwei

Computer Science. 2023, 50 (5): 155-160. doi:10.11896/jsjkx.220400035

Abstract

PDF(4328KB) ( 559 )

References | Related Articles | Metrics

Convolutional neural networks(CNNs) are widely seen in in hyperspectral image classification due to their remarkably good local context modeling performance.However,under its inherent limitations of network structure,it fails to exploit and represent sequence attributes from spectral characteristics.To address this problem,an integrated novel network,based on Swin Transformer and 3D residual multi-layer fusion network model,is proposed for hyperspectral image classification.In order to excavate the deep features of hyperspectral images as much as possible,spatial spectrum is extracted by improved 3D residual multi-layer fusion network in ReSTrans network,and the context information in consecutive spectra is captured by self-attention mecha-nism Swin Transformer network model.The final result of classification is obtained by multi-layer perception based on spatial spectrum joint feature.In order to verify the effectiveness of the ReSTrans network model,the improved model is experimentally verified on three hyperspectral data sets of IP,UP and KSC,and the classification accuracy reaches 98.65 %,99.64% and 99.78% respectively.Compared with SST method,the classification performance of the network model improves by 3.55%,0.68% and 1.87% respectively.Experimental results show that the model had good generalization ability and could extract deeper and discriminative features.

Land Use Multi-classification Method of High Resolution Remote Sensing Images Based on MLUM-Net

HU Shaokai, HE Xiaohui, TIAN Zhihui

Computer Science. 2023, 50 (5): 161-169. doi:10.11896/jsjkx.220300110

Abstract

PDF(5205KB) ( 507 )

References | Related Articles | Metrics

Aiming at the problems of incomplete land plot structure and poor boundary quality in high-resolution remote sensing image land use multi-classification,a multi-classification method of remote sensing image land use based on MLUM-Net model is proposed.This method uses the multi-scale hole convolution and channel attention mechanism to construct the MDSPA encoder,which improves the network multi-scale feature extraction ability and the accuracy of the parcel's location and adaptively enhances the multi-scale feature expression through the spatial attention mechanism.To eliminate the semantic loss of upsampling and reduce the noise of classification results,a hybrid pooling upsampling optimization module is designed to optimize the classification results and eliminate the network classification errors.According to the characteristics of unbalanced classification ratio of multi-classification data set of land use and the similarity index of plot structure,this paper designs a mixed loss function to eliminate the influence of data category ratio.This function improves the structural integrity of the block and refines the classification boundary.Experimental verification has been carried out on multiple data sets,and the overall accuracy and kappa index have been significantly improved.The classification result has a complete structure and accurate edge division,which has good practical value in the land use multi-classification.

SSD Object Detection Algorithm with Residual Learning and Cyclic Attention

JIA Tianhao, PENG Li

Computer Science. 2023, 50 (5): 170-176. doi:10.11896/jsjkx.220400085

Abstract

PDF(2244KB) ( 470 )

References | Related Articles | Metrics

To address the problem that the shallow feature semantic information generated in the feature pyramid of Single-Shot Detection is insufficient,resulting in poor performance of small object detection,an SSD object detection algorithm based on resi-dual learning with cyclic attention is proposed.Firstly,the backbone network uses Resnet101,which is more capable of learning,to extract valid feature information.The deep feature layer of the original feature pyramid is then fused with the shallow feature layer by constructing a lightweight one-way feature fusion block,and a new feature pyramid is generated,which in turn enriches the semantic information of the effective feature layer used for prediction.Finally,a new spatial pooling strategy is proposed and combined with jump connections in residual networks to form a cyclic attention module to introduce global contextual information and establish full image dependencies for local features.To address the imbalance in the number of difficult and easy samples,Focalloss is used as the regression loss function.Experimental results show that the average detection accuracy(mAP) of the algorithm is 79.7% on the PASCAL VOC public dataset,an improvement of 2.5 % over SSD.The mAP on the MS COCO public dataset is 30.0%,an improvement of 4.9 % over SSD.

Survey of Visual Question Answering Based on Deep Learning

LI Xiang, FAN Zhiguang, LI Xuexiang, ZHANG Weixing, YANG Cong, CAO Yangjie

Computer Science. 2023, 50 (5): 177-188. doi:10.11896/jsjkx.220500124

Abstract

PDF(2834KB) ( 784 )

References | Related Articles | Metrics

Visual question answering(VQA) is an interdisciplinary research paradigm that involves computer vision and natural language processing.VQA generally requires both image and text data to be encoded,their mappings learned,and their features fused,before finally generating an appropriate answer.Image understanding and result reasoning are therefore vital to the performance of VQA.With its importance in realizing cross-modal human-computer interaction and its promising applications,a number of emerging techniques for VQA,including scene-reasoning based methods,contrastive-learning based methods,and 3D-point-cloud based methods,have been recently proposed.These methods,while achieving notable performances,have revealed issues such as insufficient inferential capability and interpretability,which demand further exploration.We hence present in this paper an in-depth survey and summary of related research and proposals in the field of VQA.The essential background of VQA is first introduced,followed by the analysis and summarization of state-of-art approaches and datasets.Last but not least,with the insight of current issues,future research directions in the field of VQA are prospected.

Review of Document-level Relation Extraction Techniques

ZHU Taojie, LU Jicang, ZHOU Gang, DING Xiaoyao, WANG Ling, ZHU Xiubao

Computer Science. 2023, 50 (5): 189-200. doi:10.11896/jsjkx.220400252

Abstract

PDF(2583KB) ( 532 )

References | Related Articles | Metrics

Relation extraction(RE) is an essential direction of information extraction research,it gradually expanding from sentence to document-level.Compared with sentences,documents usually contain more relation facts,providing more information for knowledge base construction,information retrieval,and semantic analysis.However,document-level relation extraction is more complex and challenging,and there is currently a lack of systematic and comprehensive sorting and summary.To better promote the development of document-level relation extraction,this paper carries out a comprehensive and in-depth analysis of the existing technologies and methods.From the perspective of data preprocessing methods and core algorithms,it classifies the existing methods into three types,including tree-based,sequence-based,and graph-based.Based on this,Relation extraction by category analyzes and describes some typical methods,the latest progress and shortcomings.At the same time,it introduces some corpus,evaluation metrics and some typical methods.Finally,the existing problems in document-level relation extraction research are analyzed and summarized,and the possible future development trends and research directions are discussed.

Survey on Knowledge Transfer Method in Deep Reinforcement Learning

ZHANG Qiyang, CHEN Xiliang, CAO Lei, LAI Jun, SHENG Lei

Computer Science. 2023, 50 (5): 201-216. doi:10.11896/jsjkx.220400235

Abstract

PDF(3352KB) ( 912 )

References | Related Articles | Metrics

Deep reinforcement learning is a hot issue in artificial intelligence research.With the deepening of research,some shortcomings are gradually exposed,such as low data utilization,weak generalization ability,difficult exploration,lack of reasoning and representation ability,etc.These problems greatly restrict the application of deep reinforcement learning method in practical pro-blems.Knowledge transfer is a very effective method to solve this problem.This study discusses how to use knowledge transfer to accelerate the process of agent training and cross domain transfer from the perspective of deep reinforcement learning,analyzes the existing forms and action modes of knowledge in deep reinforcement learning,and classifies and summarizes the knowledge transfer methods in deep reinforcement learning according to the basic elements of reinforcement learning.Finally,the existing problems and cutting-edge development direction of knowledge transfer in deep reinforcement learning in algorithm,theory and application are reported.

Overview of Intelligent Radar Fault Prediction and Detection Technology

ZHAI Yuting, CHENG Zhanxin, FANG Shaojun

Computer Science. 2023, 50 (5): 217-229. doi:10.11896/jsjkx.220400096

Abstract

PDF(2137KB) ( 723 )

References | Related Articles | Metrics

Radar fault prediction and fault detection technology is the key technology for the transformation of radar equipment maintenance from traditional regular maintenance to intelligent condition-based maintenance.To ensure the performance of radar combat effectiveness,it is necessary to predict,detect and give real-time warnings to radar fault in time.With the maturity of microwave measurement and artificial intelligence technology,intelligent radar fault prediction and detection technology also conti-nues to develop.In this paper,the current research status of fault prediction and health management and fault detection technology at home and abroad are elaborated,the advantages and disadvantages of the existing intelligent radar fault prediction and detection technology are analyzed,the research progress of this technology in the field of radar maintenance support has been sorted out,possible problems and limitations in radar failure prediction and detection are presented.Aiming at the actual problems and constraints,the future research direction of intelligent radar fault prediction and detection technology are prospected,which can provide a reference for the in-depth research of intelligent fault prediction and fault detection technology in the field of radar maintenance support.

Aspect-based Sentiment Analysis Based on Dual-channel Graph Convolutional Network with Sentiment Knowledge

YANG Ying, ZHANG Fan, LI Tianrui

Computer Science. 2023, 50 (5): 230-237. doi:10.11896/jsjkx.220300008

Abstract

PDF(2066KB) ( 477 )

References | Related Articles | Metrics

Aspect-based sentiment analysis is a fine-grained sentiment analysis task whose goal is to classify the sentiment polarity of the given aspect terms in a sentence.Most of the current sentiment classification models build a graph neural network on the dependency syntax tree,and learn the information between the aspect terms and the context from the dependency syntax tree,and lack the mining of sentiment knowledge in the sentence.To solve this problem,this paper proposes a sentiment classification model based on dual-channel graph convolutional network with sentiment knowledge.The model consists of a sentiment-enhanced dependency graph convolutional network(SDGCN) and an attention graph convolutional network(AGCN),which learn the syntactic dependencies and semantic relations of aspect terms and context words,respectively.Specifically,SDGCN incorporates sentiment knowledge from SenticNet on syntactic dependencies to enhance sentence dependencies,so that the model considers the syntactic relationship between context and aspects,together with the sentiment information between opinion words in the context and aspect terms.The attention mechanism is used by AGCN to learn the semantic relevance between aspect terms and the context in the sentence.Finally,the two graph convolution networks learn their own information interactively for sentiment classification.Experimental results show that the proposed model performs well on multiple public datasets,and ablation experiments verify the effectiveness of each module.

Sentiment Analysis Based on Multi-event Semantic Enhancement

ZHANG Xue, ZHAO Hui

Computer Science. 2023, 50 (5): 238-247. doi:10.11896/jsjkx.220400256

Abstract

PDF(2803KB) ( 486 )

References | Related Articles | Metrics

Implicit sentiment analysis is to detect the sentiment of sentences that do not contain obvious sentiment words.This paper focuses on event-centric sentiment analysis,which is to infer sentiment polarity from the events described in the sentence.In event-centric sentiment analysis,existing methods either treat noun phrases in the text as events,or adopt complex models to model events,fail to model event information sufficiently,and fail to consider events that contain multiple events.In order to solve the above problems,it is proposed to represent events in the form of event triples〈subject,predicate,object〉.Based on this event representation,an event-enhanced semantic-based sentiment analysis model(MEA) is further proposed to detect the sentiment of texts.In this paper,syntactic information is used to capture the relationship of event triples,and attention mechanism is used to model the relationship between events according to the contribution of each event to the sentence.At the same time,a bidirec-tional long-short-term memory network(Bi-LSTM) is used to model the contextual information of sentences,and a multi-level orthogonal attention mechanism is used to capture the difference of attention weights under different polarities,which can be used as a significant discriminative feature.Finally,according to the importance of event features and sentence features,they are assigned different weight ratios,and they are fused to obtain the final sentence representation.Furthermore,this paper proposes a dataset for event-enhanced sentiment analysis(MEDS),where each sentence is labeled with event triplet representations and sentiment polarity labels.Research shows that the proposed model outperforms existing baseline models in self-built datasets.

Chinese Sentiment Analysis Based on CNN-BiLSTM Model of Multi-level and Multi-scale Feature Extraction

WANG Lin, MENG Zuqiang, YANG Lina

Computer Science. 2023, 50 (5): 248-254. doi:10.11896/jsjkx.220400069

Abstract

PDF(2278KB) ( 751 )

References | Related Articles | Metrics

Sentiment analysis,as a sub-field of natural language processing(NLP),plays a very important role in public opinion monitoring.In the Chinese sentiment analysis task,the existing methods only consider sentiment features from single-level and single-scale,which cannot fully mine and utilize the sentiment feature information,and the performance of the model is not ideal.To solve this problem,a CNN-BiLSTM model with multi-level and multi-scale feature extraction is proposed.This model first uses a pre-trained Chinese word vector model combined with embedding layer fine-tuning to obtain word-level features.Then,phrase-level and sentence-level features are obtained by multi-scale phrase-level feature representation module and sentence-level feature representation module respectively.In the multi-scale phrase-level feature representation module,convolutional networks with different convolution kernel sizes are used to obtain phrase-level features of different scales.Finally,a multi-level feature fusion method is used to fuse word-level features,phrase-level features of different scales,and sentence-level features to form multi-level joint features.Compared with single-level and single-scale features,multi-level joint features have more sentiment information.In the experiment,four evaluation indicators(Accuracy,Precision,Recall,F1) are used to evaluate the performance of the model and compared with eight methods including support vector machines(SVM).Experimental results show that the proposed method outperforms the eight comparison methods in the four evaluation indicators,which proves the advantages of the proposed model in multi-level and multi-scale feature extraction.

Document-level Event Extraction Based on Multi-granularity Entity Heterogeneous Graph

ZHANG Hu, ZHANG Guangjun

Computer Science. 2023, 50 (5): 255-261. doi:10.11896/jsjkx.220300154

Abstract

PDF(3981KB) ( 422 )

References | Related Articles | Metrics

Document-level event extraction is an event extraction task for long texts with multiple sentences.Existing document-level event extraction studies generally divide event extraction into three sub-tasks:candidate entity extraction,event detection and argument recognition,and usually train them with joint learning.However,most of the existing document-level event extraction methods extract candidate entities sentence-by-sentence without considering the cross-sentence contextual information,which obviously reduces the accuracy of entity extraction and argument recognition.Furthermore,it affects the final event extraction results.Based on this,this paper proposes a document-level event extraction method based on multi-granularity entity heteroge- neous graphs.This method uses two independent encoders,Transformer and RoBerta,for sentence-level and paragraph-level entity extraction respectively.Meanwhile,this paper proposes a multi-granularity entity fusion strategy to select entities that are more likely to be event arguments from the set of sentence entities and paragraph entities,and further constructs a heterogeneous graph incorporating multi-granularity entities.Finally,we use graph convolutional network to obtain document-aware entity and sentence representations for multi-label classification of event types and event arguments to achieve event detection and arguments recognition.Experiments on ChFinAnn and Duee-fin datasets show that the proposed method improves about 1.3% and 3.9% by F1 compared with previous methods,which proves its effectiveness.

Convolutional Network Entity Missing Detection Method Combined with Gated Mechanism

YE Han, LI Xin, SUN Haichun

Computer Science. 2023, 50 (5): 262-269. doi:10.11896/jsjkx.220400126

Abstract

PDF(2491KB) ( 408 )

References | Related Articles | Metrics

The adequacy of the entity information directly affects the applications that depend on textual entity information,while conventional entity recognition models can only identify the existing entities.The task of the entity missing detection,defined as a sequence labeling task,aims at finding the location where the entity is missing.In order to construct training dataset,three corres-ponding methods are proposed.We introduce an entity missing detection method combining the convolutional neural network with the gated mechanism and the pre-trained language model.Experiments show that the F1 scores of this model are 80.45% for the PER entity,83.02% for the ORG entity,and 86.75% for the LOC entity.The model performance exceeds the other LSTM-based named entity recognition model.It is found that there is a correlation between the accuracy of the model and the word frequency of the annotated characters.

Answer Selection Model Based on MLP and Semantic Matrix

LUO Liang, CHENG Chunling, LIU Qian, GUI Yaocheng

Computer Science. 2023, 50 (5): 270-276. doi:10.11896/jsjkx.220400275

Abstract

PDF(1790KB) ( 413 )

References | Related Articles | Metrics

Answer selection is a key sub-task in the field of question answering systems,and its performance supports the deve-lopment of question answering systems.The dynamic word vector generated by the BERT model based on parameter freezing also has problems such as lack of sentence-level semantic features and the lack of word-level interaction between question and answer.Multilayer perceptrons have a variety of advantages,they not only can achieve deep feature mining,but also have low computational costs.On the basis of dynamic text vectors,this paper proposes an answer selection model based on multi-layer perceptrons and semantic matrix,which mainly realizes the semantic dimension reconstruction of text vector sentences,and generates semantic matrix through different calculation methods to mine different text feature information.The multi-layer perceptron is combined with the semantic understanding matrix generated by the linear model to implement a semantic understanding module,which aims to excavate the sentence-level semantic characteristics of the question sentence and the answer sentence respectively; the multi-layer perceptron is combined with the semantic interaction matrix generated based on the two-way attention calculation method to achieve a semantic interaction module,which aims to build the word-level interaction relationship between the question and answer pairs.Experimental results show that the proposed model has a MAP and MRR of 0.789 and 0.806 on the WikiQA dataset,respectively,which has a consistent performance improvement over the baseline model,on the SelQA dataset,MAP and MRR is 0.903 and 0.911,respectively,which also has a good performance.

Binary Harris Hawk Optimization and Its Feature Selection Algorithm

SUN Lin, LI Mengmeng, XU Jiucheng

Computer Science. 2023, 50 (5): 277-291. doi:10.11896/jsjkx.220300269

Abstract

PDF(2802KB) ( 635 )

References | Related Articles | Metrics

Harris Hawk optimization(HHO) algorithm only uses the random strategy to initialize the population in the exploration stage,which decreases the population diversity.The escape energy that controls the linear variation of the development and exploration process is prone to fall into local optimum in the later stage of iteration.To address the issues,this paper proposes a binary Harris Hawk optimization for metaheuristic feature selection algorithm.First,in the exploration phase,the Sine mapping function is introduced to initialize the population location of Harris Hawk,and the adaptive adjustment operator is used to change the search range of HHO and update the population location of HHO.Second,the updated formula of escape energy is improved by the logarithmic inertia weight,the number of iterations are introduced into the jump distance,and the step size adjustment parameter is employed to adjust the search distance of HHO to balance the exploration and development capabilities.On this basis,an improved HHO algorithm is designed to avoid the HHO algorithm falling into the local optimum.Third,the binary position and population position of the improved HHO algorithm are updated by the S-type and V-type transfer functions.Thus two binary improved HHO algorithms are designed.Finally,a fitness function is used to evaluate the feature subset,the binary improved HHO algorithm is combined with this fitness function,and then two binary improved HHO metaheuristic feature selection algorithms are developed.Experimental results on 10 benchmark functions and 17 public datasets show that the four optimization strategies effectively improve the optimization performance of the HHO algorithms on these benchmark functions,and the improved HHO algorithm is significantly better than other compared optimization algorithms.On 12 UCI datasets and 5 high-dimensional gene datasets,when compared with the BHHO-based feature selection algorithms and the other feature selection algorithms,the results demonstrate that the V-shape-based improved HHO feature selection algorithm has great optimization ability and classification performance.

Coke Price Prediction Based on ELM Optimized by Double-elite Evolution Salp Swarm Algorithm

ZHU Xuhui, SHE Xiaomin, NI Zhiwei, XIA Pingfan, ZHANG Chen

Computer Science. 2023, 50 (5): 292-301. doi:10.11896/jsjkx.220300259

Abstract

PDF(2811KB) ( 393 )

References | Related Articles | Metrics

Coke is one of important industrial raw materials,and accurate prediction of its future price trend has great significance for making production scheduling plans of coking plants.Extreme learning machine(ELM) has strong generalization ability and fast computing speed,and it is suitable as the model of coke price prediction.However,the prediction performance of ELM is greatly affected by its key parameters,and its parameters need to be optimized.Based on this,a coke price prediction method is proposed by optimizing the key parameters of ELM using double-elite evolution salp swarm algorithm.Firstly,the double-elite evolutionary salp swarm algorithm(MDSSA) is proposed by introducing logistic chaotic mapping,improved convergence factor,adaptive inertia weights and double-elite evolutionary mechanism,so as to enhance the search capability of salp swarm algorithm(MDSSA).Secondly,the connection weights and thresholds of ELM are optimized using MDSSA for finding the optimal parameters combination,so as to construct the MDSSA-ELM coke price prediction model.Finally,the convergence performance of MDSSA is validated using 8 benchmark functions,and the prediction ability of MDSSA-ELM model is tested on the actual coke price dataset.Experimental results demonstrate that MDSSA-ELM has stronger predictive capability than other methods,and MDSSA has superior searching ability than other algorithms,which provides an effective prediction tool for coking plants for achieving intelligent production scheduling.

Task Offloading Strategy Based on Game Theory in 6G Overlapping Area

GAO Lixue, CHEN Xin, YIN Bo

Computer Science. 2023, 50 (5): 302-312. doi:10.11896/jsjkx.220500120

Abstract

PDF(2871KB) ( 487 )

References | Related Articles | Metrics

In order to realize the efficient computing of complex tasks in the overlapping area of 6G network base station(BS) service,the task offloading problem in the overlapping area is studied.Based on the comprehensive consideration of delay constraints,energy consumption,social effects and economic incentives,a multi-access edge computing network model with multiple BSs and multiple Internet of things(IoT) devices is constructed,and the BSs pricing strategy,the base station selection strategy and the task offloading strategy of IoT devices are jointly optimized to maximize the profit of BSs and the utility of IoT devices.To solve the problem of base station selection for IoT devices in overlapping areas,a many-to-one matching game model is built,and the BSs selection algorithm based on swap matching is proposed.A two-stage game model for pricing and task offloading interaction between BSs and IoT devices is established by introducing Stackelberg game theory,the existence and uniqueness of Stackelberg equilibrium are proved by backward induction.The optimal price and best response algorithm based on game theory(OBGT) based on game theory is proposed.Simulation and comparison experiments illustrate that OBGT algorithm can achieve convergence in a short time,and effectively improve the profit of BSs and the utility of IoT devices.

Study on Load Balancing Algorithm of Microservices Based on Machine Learning

YANG Qianlong, JIANG Lingyun

Computer Science. 2023, 50 (5): 313-321. doi:10.11896/jsjkx.220400019

Abstract

PDF(3658KB) ( 541 )

References | Related Articles | Metrics

With the continuous development of cloud computing technology,the microservice architecture has received more and more attention.Since it is more convenient in development and maintenance to divide large-scale applications into fine-grained single services,lots of large applications have evolved from monolithic architecture to microservice architecture.In the microservice architecture,in order to improve the availability of microservices,microservice instances are usually deployed in a cluster structure.Aiming at the problem of unbalanced load of server nodes in a microservice cluster with the increase of the number of tasks,a load balancing algorithm based on Xgboost,shortest predictive response time,is proposed.By selecting the characteristic para-meters that affect the response time of the task,and then using machine learning to predict the response time of new task,the task is finally assigned to the server node with the smallest predicted response time,so as to achieve the purpose of load balancing between server nodes.The results show that using the proposed load balancing algorithm has a certain improvement in throughput,cut-off rate and average response time compared with other load algorithms,and it is more suitable for microservice clusters in high concurrency environments.

Anti-interference Multiuser Detection Algorithm Based on Variable Step Size Adaptive Matching Pursuit in Grant-free NOMA System

LI Yuge, WANG Tianjing, SHEN Hang, LUO Xiaokang, BAI Guangwei

Computer Science. 2023, 50 (5): 322-328. doi:10.11896/jsjkx.220400170

Abstract

PDF(2655KB) ( 434 )

References | Related Articles | Metrics

The fifth generation mobile communication system(5G) uses non-orthogonal multiple access(NOMA) technology for non-orthogonal multiplexing of wireless communication resources,which improves the spectrum utilization efficiency and system capacity by the way of overload.The NOMA system uses the grant-free mode to reduce the system flow and signaling overhead,but the receiver needs to perform multi-user detection.Based on the sparse characteristics of active users,the base station uses the compressed sensing(CS) reconstruction algorithm to recover the mixed sparse vectors of active users,and realizes efficient multi-user detection.The dense deployment of base stations in 5G network enhances the interferences among neighboring cells that increases the difficulty of CS-based detection and reduces the accuracy of detection.Aiming at the problem of interference in multi-user detection in the grant-free NOMA system,an anti-interference multiuser detection algorithm based on variable step size adaptive matching pursuit is proposed.Unknowing the sparse degree,the anti-interference active user detection can be realized by the adaptive variable step size way,in which the sparse degree is fast approached with large step size and accurately approximated with small step size.Simulation results show that,under different overload rates,the bit error rates of the proposed algorithm are lower than that of traditional multi-user detection algorithms based on OMP,gOMP and SAMP.

Review of Identity Authentication Research Based on Blockchain Technology

ZHANG Shue, TIAN Chengwei, LI Baogang

Computer Science. 2023, 50 (5): 329-347. doi:10.11896/jsjkx.220400169

Abstract

PDF(3421KB) ( 1195 )

References | Related Articles | Metrics

Blockchain technology was proposed by Nakamoto in his 2008 white paper.As a decentralized and distributed public ledger technology in point-to-point networks,blockchain verifies and stores data by applying link block structure,and synchro-nizes data changes by applying trusted consensus mechanism,providing a trusted technical solution for the realization of identity authentication.Compared with traditional centralized authentication,identity authentication based on blockchain technology can realize data sharing while protecting the authenticity and reliability of data and the privacy and security of nodes.This paper summarizes the status and progress of identity authentication based on blockchain technology.Firstly,it systematically introduces some basic theories of blockchain from the technical architecture,classification and consensus algorithm of blockchain.Next,it focuses on password authentication technology,biometric technology,PKI technology and the current research status of identity authentication combined with blockchain application.Then it introduces the research progress of identity authentication technology based on blockchain from the application fields of Internet of things,Internet of vehicles,smart grid,finance,medical treatment and so on.Finally,it analyzes the current problems of blockchain identity authentication technology,and puts forward the future development trend.

Double Dummy Location Selection Algorithm Based on Behavior Correlation

TU Sipan, ZHANG Lin, LIU Xiping

Computer Science. 2023, 50 (5): 348-354. doi:10.11896/jsjkx.220300207

Abstract

PDF(2511KB) ( 435 )

References | Related Articles | Metrics

At present,most of the dummy location privacy protection methods do not fully consider that the attacker may have the background knowledge about the information,which user requested from the service last time,so it cannot effectively resist the background knowledge attack.Based on this,a double dummy location generation algorithm based on behavior correlation is proposed,which consists of two parts.First,a first dummy location generation algorithm is studied by taking full account of the access probability and location semantics.Based on the user behavior information such as time accessibility and direction similarity,a regeneration dummy location algorithm is given.Finally,a large number of simulation experiments verify that this algorithm can resist inference attack and similarity attack.

Multimodal Generative Adversarial Networks Based Multivariate Time Series Anomaly Detection

ZHANG Renbin, ZUO Yicong, ZHOU Zelin, WANG Long, CUI Yuhang

Computer Science. 2023, 50 (5): 355-362. doi:10.11896/jsjkx.220400221

Abstract

PDF(2110KB) ( 570 )

References | Related Articles | Metrics

Aiming at the problem that the traditional anomaly detection model of multivariate time series data does not consider the multimodal distribution of spatio-temporal data,a multivariate time series data anomaly detection model based on multimodal generative adversarial networks is proposed.The sliding windows is used to segment the time series and construct feature matrices,so as to capture the multimodal features of the data.Feature matrix and raw data are fed into the multimodal encoder and multimodal generator as modal information respectively,then multimodal feature matrix with spatio-temporal information is outputted.The real data is encoded into feature matrices and the two types of feature matrices are utilized as discriminator inputs.In the proposed method,a gradient penalty method and the Wasserstein distance between the real and generated distributions to replace the binary cross-entropy loss are utilized to train the discriminator,then combining the generator reconstruction error and discriminator scores to detect anomalies.Experimental results based on the secure water treatment(SWaT) and the water distribution(WADI) datasets show that,compared with the baseline model,the proposed method improves the F1-score metrics by 0.11 and 0.19 respectively.The proposed method can identify multivariate time series data anomalies well,with good robustness and generalizability.

Provably Secure Key Management Protocol for Heterogeneous WSN

ZHANG Linghao, TANG Yong, DENG Dong, LIU Yangyang, TANG Chao, GUI Shenglin

Computer Science. 2023, 50 (5): 363-371. doi:10.11896/jsjkx.220400193

Abstract

PDF(2534KB) ( 394 )

References | Related Articles | Metrics

Authentication and key agreement protocol is the mainstream method to solve the secure communication of devices in wireless sensor networks.For the current mainstream secret key agreement protocol in wireless sensor networks,the scenario considered is authentication and agreement between peer nodes,which has the problems of high computation and low communication efficiency.To solve the above problems,this paper proposes an authentication and secret key management protocol suitable for heterogeneous wireless sensor networks.Both communication terminal nodes(L node) first establish the session secret key with the management node(H node) of their respective cluster.If they fail to pass the identity authentication,the L node will be denied access to the network,which solves the problem that most protocols lack to deal with denial of service attacks.Then,with the help of the H node through which the communication path passes,the session key agreement information is forwarded to complete the end-to-end session key agreement between the communication parties,so that the protocol has the abilities of security gateway and access control.The protocol also supports the revocation of the captured node and reduces the impact on the security of other communication links.Based on the difficult assumption of solving the discrete logarithm problem and Diffie Hellman problem on elliptic curve,it is proved in the random oracle model that the scheme can meet more complete security attributes such as forward security,anti secret key leakage camouflage attack,unknown secret key sharing security,no secret key escrow,known secret key security and so on.Compared with the existing literature,the protocol has the lowest computational overhead in dealing with denial of service attack,and the overall amount of computation and communication is moderate.

Neural Network Model Training Method Based on Homomorphic Encryption

ZHAO Min, TIAN Youliang, XIONG Jinbo, BI Renwan, XIE Hongtao

Computer Science. 2023, 50 (5): 372-381. doi:10.11896/jsjkx.220300239

Abstract

PDF(2403KB) ( 848 )

References | Related Articles | Metrics

Aiming at the problem of data privacy leakage in cloud environment and insufficient accuracy in the privacy-preserving neural network based on homomorphic encryption,a privacy-preserving neural network training scheme(PPNT) is proposed for collaborative dual cloud servers,to achieve the goal of data transmission,computing security and model parameter under the collaborative training process of dual cloud servers.Firstly,in order to avoid using polynomial approximation method to realize nonlinear functions such as exponent and comparison,and improve the calculation accuracy of nonlinear function,a series of secure computing protocols are designed based on Paillier partially homomorphic encryption technology and additive secret sharing scheme.Furthermore,corresponding secure computing protocols of full connection layer,activation layer,softmax layer and back propagation in neural network are constructed to realize PPNT based on the designed secure computing protocols.Finally,theoretical and security analysis guarantees the correctness and security of PPNT.The actual performance results show that compared with the dual server scheme--privacy protection machine learning as a service(PPMLaaS),the model accuracy of PPNT improves by 1.7%,and supports the client offline in the process of secure computing.

Multi-source Fusion Network Security Situation Awareness Model Based on Convolutional Neural Network

CHANG Liwei, LIU Xiujuan, QIAN Yuhua, GENG Haijun, LAI Yuping

Computer Science. 2023, 50 (5): 382-389. doi:10.11896/jsjkx.220400134

Abstract

PDF(2794KB) ( 766 )

References | Related Articles | Metrics

For accurately calculating security situation of the whole network,a network security situation awareness model with five core elements is elaborated,which are traffic detection,attribute extraction,decision engine,multi-source fusion and situation assessment.In the traffic detection module,the network traffic detector and the intrusion detection detector are taken as a tool to grab the basic characteristics of traffic and malicious activity characteristics respectively; in the attribute extraction module,with the aim of precisely extracting key attributes,alarm messages,alarm types and connection characteristics,which contribute to describe malicious activities,are the center of attention; in the decision engine module,the key attribute data from attribute extraction is utilized as input,and CNN as an engine is employed to identify various kinds of attacks; in the multi-source fusion module,exponential weighted D-S fusion algorithm is used to effectively integrate the output of each decision engine to improve the identification rate of attack types; in the situation assessment module,in virtue of weight coefficient theory the threat levels are quantified,the hierarchical analysis method is applied to exactly get security situation of the whole network.Experimental results show that,there is a great difference in identifying varieties of attacks for different detectors,the proposed multi-source fusion algorithm can improve the accuracy of attack identification which can reach up to 92.76%,in such accuracy index our results are better than most research achievements,and the improvement of accuracy makes a great impact on accurately calculating and intuitively reflecting security situation of the whole network by means of hierarchical analysis method.