Computer Science

Review on Performance Optimization of Ceph Distributed Storage System

ZHANG Xiao, ZHANG Si-meng, SHI Jia, DONG Cong, LI Zhan-huai

Computer Science. 2021, 48 (2): 1-12. doi:10.11896/jsjkx.201000149

Abstract

PDF(2750KB) ( 4371 )

References | Related Articles | Metrics

Ceph is a unified distributed storage system,which can provide storage services of 3 types of interfaces:block,file and object.Different from the traditional distributed storage system,it adopts the metadata management method without central node,so it has good scalability and linear growth performance.After more than ten years of development,Ceph has been widely used in cloud computing and big data storage systems.As the underlying platform of cloud computing,Ceph not only provides storage service for virtual machines,but also directly provides the object storage service and NAS file service.Ceph supports storage requirements of various operating systems and applications in cloud computing systems.Its performance has a great influence on virtual machines and applications running on it.Therefore,the performance optimization of the Ceph storage system has been a research hotspot in academia and industry.This paper first introduces the architecture and characteristics of Ceph,then summarizes existing performance optimization technologies from 3 aspects,including internal mechanism improvement,new hardware-orien-ted and application-based optimization and reviews the recent research on Ceph storage and optimization.Finally,it prospects the future work,hoping to provide a valuable reference for researchers in the performance optimization of distributed storage system.

Reference Model and Development Methodology for Enterprise Cloud Service Architecture

JIANG Hui-min, JIANG Zhe-yuan

Computer Science. 2021, 48 (2): 13-22. doi:10.11896/jsjkx.200300044

Abstract

PDF(4541KB) ( 1462 )

References | Related Articles | Metrics

Service-oriented software engineering with the fusion of the services and cloud computing paradigms not only offers many advantages for large-scale distributed software development and applications,but also brings new challenges.The biggest challenge in cloud computing is the lack of a de facto standard or single architectural design method,which can meet the requirements of an enterprise cloud approach to help deliver software as a service over the Internet.First,according to the business cha-racteristics of enterprise cloud computing,a generic and abstract model for Enterprise Cloud Service Architecture (ECSA) is proposed.The model consists of nine components,including the cloud services,service mode,service consumers,management,processes,quality attributes,service matching and interactive matching.The model components and their relationships are analysed,and their roles are discussed.Then,a four-phase software architecture improvement process that considers cloud services as the first class modeling elements is also presented.By decoupling the cloud service mode from their implementation on target component configurations,the process supports exploration of multiple architectures utilizing the same set of services.Finally,the application instance of ECSA is introduced,which hopes to provide recommendations and reference for enterprise cloud service system development and application integration.

Collaborative Scheduling of Source-Grid-Load-Storage with Distributed State Awareness UnderPower Internet of Things

WANG Xi-long, LI Xin, QIN Xiao-lin

Computer Science. 2021, 48 (2): 23-32. doi:10.11896/jsjkx.200900209

Abstract

PDF(2504KB) ( 1366 )

References | Related Articles | Metrics

With the development of new generation,direct-current transmission,electric energy storage and other technologies,flexible load such as new energy generation and electric vehicles and energy storage devices with charge-discharge ability are constantly integrated into the power grid,which makes the traditional distribution network architecture change greatly.Due to the great instability of the new type of source grid load storage,it brings great challenges to the distribution network dispatching,especially the extra power loss in scheduling which is difficult to control.With the construction of Ubiquitous Power Internet of Things (UPIoT),real-time information collection and data analysis of source grid load storage can be realized,which provides an opportunity for real-time data-driven collaborative scheduling of Source-Grid-Load-Storage.The collaborative scheduling of Source-Grid-Load-Storage in distribution network has a natural distributed characteristic.Therefore,a distributed state awareness system can be built which can bring low latency and high precision for the collaborative real-time scheduling of Source-Grid-Load-Storage.The distribution network structure under the background of UPIoT is analyzed in this paper,then the source grid load storage and their interaction methods in a distributed environment are modeled.This model is based on the premise that the feeder nodes have certain computing and communication capabilities,and it stipulates the data interaction method of all the nodes in entire distribution network,which can effectively reflect the effect of collaborative scheduling in the distribution network.A collaborative scheduling mechanism of Source-Grid-Load-Storage with distributed state awareness under Power Internet of Things is proposed,and the response strategy of each end of source grid load storage is defined in this paper,thus realizing the goal of peak load shifting and scheduling loss reduction.Based on some real data of the power grid,a simulation verification experiment is designed.The experimental results verify the effectiveness of the collaborative scheduling mechanism of Source-Grid-Load-Storage.

Method of Encapsulating Procuratorate Affair Services Based on Microservices

LU Yi-fan, CAO Rui-hao, WANG Jun-li, YAN Chun-gang

Computer Science. 2021, 48 (2): 33-40. doi:10.11896/jsjkx.191100152

Abstract

PDF(1937KB) ( 1009 )

References | Related Articles | Metrics

Microservice architecture is an emerging style of service architecture,which is characterized by efficient operation and flexible deployment when dealing with complex service systems.Compared with monolithic architecture,it can provide better business management and service support.In view of the complex case of the procuratorate affair,it is necessary to combine and encapsulate the services to form new value-added services to meet the needs of users.However,quality-of-service driven service encapsulation alone cannot meet the needs of procuratorate affair.Therefore,combining service functions and quality of service,an improved graphplan under microservice architecture (IGMA) is proposed.Firstly,the method establishes a mathematical model for the service and user request,then integrates the functional and non-functional requirements of the service,and provides users with a variety of combination schemes under different case types.Finally,the service workflow is established to complete the case service encapsulation.This method can intelligently judge the branch structures in the service composition structure and establish different composition schemes for different branch structures.Experimental results show that the proposed method improves the timeliness and accuracy of service encapsulation.

Intermediate Data Transmission Pipeline Optimization Mechanism for MapReduce Framework

ZHANG Yuan-ming, YU Jia-rui, JIANG Jian-bo, LU Jia-wei, XIAO Gang

Computer Science. 2021, 48 (2): 41-46. doi:10.11896/jsjkx.191000103

Abstract

PDF(2370KB) ( 843 )

References | Related Articles | Metrics

MapReduce is an important parallel computing framework for large data processing,which greatly improves the performance of data processing by performing multiple tasks in parallel on a large number of cluster nodes.However,since the intermediate data needs to wait until the Mapper task is completed,it can be sent to the Reducer task.The massive transmission delay becomes an important bottleneck of the MapReduce framework performance.To this end,an intermediate data transmission pipeline mechanism for MapReduce is proposed.It decouples the effective computation from intermediate data transmission,overlaps each stage in pipeline mode,and effectively hides data transmission delay.The execution mechanism and implementation strategy of the approach are given,including pipeline partition,data subdivision,data merging and data transmission granularity.The proposed mechanism is evaluated on public data sets.When the Shuffle data volume is large,the overall performance improves by 60.2% compared with the default framework.

Convolutional Optimization Algorithm Based on Distributed Coding

YUAN Chen-yu, XIE Zai-peng, ZHU Xiao-rui, QU Zhi-hao, XU Yuan-yuan

Computer Science. 2021, 48 (2): 47-54. doi:10.11896/jsjkx.200800187

Abstract

PDF(1933KB) ( 1039 )

References | Related Articles | Metrics

Convolution operation plays a vital role in statistics,signal processing,image processing and deep learning.It is also an important operation in deepneural networks where it is the basis of information filter and characteristics extraction.The exploration of methods to speed up the convolutional operations has become an open research topic in recent years.Many studies have pointed out that distributed computing framework may improve the computational time of convolution operations and hence optimize the training efficiency for deep learning.But stragglers in distributed systems may slow down the overall system.This paper proposes a distributed coding based Winograd algorithm for implementing 2D convolution operations.In this algorithm,the Winograd algorithm can effectively accelerate the speed of 2D convolution calculation,while distributed encoding can mitigate the impact of stragglers by using a redundancy-based encoding strategy.Therefore,the proposed distributed 2D convolution algorithm can effectively mitigate the straggler problem in distributed systems while improving the 2D convolution calculation,hence it may effectively improve the computational efficiency of distributed 2D convolution algorithms.

Study on Heterogeneous UAV Formation Defense and Evaluation Strategy

ZUO Jian-kai, WU Jie-hong, CHEN Jia-tong, LIU Ze-yuan, LI Zhong-zhi

Computer Science. 2021, 48 (2): 55-63. doi:10.11896/jsjkx.191100053

Abstract

PDF(3952KB) ( 1022 )

References | Related Articles | Metrics

The problem of UAV formation confrontation has always been a hot topic in scientific research,and there are few related studies on the deployment of UAV group defense.Based on the protection of defensive UAV against common UAV,such as civil,commercial,reconnaissance,cruise and exploration,the coding and decoding scheme of existing heterogeneous UAV formation is improved.The fitness function is established from the missile flight distance and the safety of unarmed drones,and the genetic algorithm is used to optimize the defense formation of the drone.According to the situation of enemy UAV of different sizes and various formations,the formation of our UAV is optimized.The solution results show that the genetic algorithm can converge to the optimal value in different enemy formations at a high speed within 30 iterations,and the corresponding optimized formation is given.Finally,by evaluating the probability effect and drawing the loss curve of five combat situations,it can be seen that the defense deployment strategy designed in this paper is effective.The maximum loss quantity of our UAVs is 6,minimum loss quantity is 0,average loss quantity is 3,and average loss rate is 18.75%.This method is of great significance for the research of UAV group defense deployment.

Fine-grained Performance Analysis of Uplink in Wireless Relay Network Based on Stochastic Geometry

SUN Hai-hua, ZHOU Si-yuan, TAN Guo-ping, ZHANG Zhi

Computer Science. 2021, 48 (2): 64-69. doi:10.11896/jsjkx.200800205

Abstract

PDF(2174KB) ( 769 )

References | Related Articles | Metrics

As the number of wireless network users increases dramatically,the network topology of traditional cellular networks can't meet the performance requirements of all users.In order to improve the uplink coverage probability in the cell-edge area,an uplink amplify-and-forward (AF) relay network model in which relays are deployed around the base station is established.The base stations and relays are respectively modeled as the Poisson point process (PPP) and the truncated Thomas cluster process (TCP),and relay node amplifies and forwards the data of cell-edge users to the base station.We drive a fine-grained performance analysis of the network model,i.e.,SIR meta distribution which is the distribution of the conditional coverage probability (CCP).The moments of the CCP in the relays network are analytically derived and the approximation of SIR meta distribution is presented in semiclosed-form expression.In contrast to the conventional performance analysis based on the coverage probability,meta distribution can intuitively show the proportion of uplinks in the network whose CCP is greater than a certain value.And the accuracy of the theoretical analysis is verified by simulations.Besides,the effect of the relay distribution parameters on the meta distribution is studied by adjusting the parameters of the relay distribution,such as radius and variance.Finally,the effects of power compensation factor of the uplink power control on the network coverage probability are compared,which provides help for the later research on network performance optimization.

SpaRC Algorithm Hyperparameter Optimization Methodology Based on TPE

DENG Li, WU Jin-da, LI Ke-xue, LU Ya-kang

Computer Science. 2021, 48 (2): 70-75. doi:10.11896/jsjkx.200500156

Abstract

PDF(2102KB) ( 749 )

References | Related Articles | Metrics

The assembly of metagenomic sequences faces huge challenge in computing and storage.SpaRC (Spark Reads Clustering) is a metagenomic sequence fragment clustering algorithm based on Apache Spark,which provides a scalable solution for clustering of billions of sequencing fragments.However,setting SpaRC parameters is a very challenging task.SpaRC algorithm has many hyperparameters that have a great impact on the performance of the algorithm.Choosing the appropriate hyperparameter set is crucial to the performance of SpaRC algorithm.In order to improve the performance of SpaRC algorithm,a hyperpara-meter optimization method based on Tree Parzen Estimator (TPE) is explored,which can use prior knowledge to efficiently adjust the parameters,accelerate the search for the optimal parameters by reducing the calculation task to achieve the optimal clustering effect,thus avoding expensive parameter exploration.After experiments with long-reads(PacBio) and short-reads(CAMI2),the results show that the proposed method has a great effect on improving the performance of SpaRC algorithm.

Modeling Methods of Social Network User Influence

TAN Qi, ZHANG Feng-li, ZHANG Zhi-yang, CHEN Xue-qin

Computer Science. 2021, 48 (2): 76-86. doi:10.11896/jsjkx.191200102

Abstract

PDF(2085KB) ( 2734 )

References | Related Articles | Metrics

Social network user influence has been widely used in the fields of public opinion evolution,advertising marketing,political election,etc.In the past work,researchers have achieved certain results by analyzing and modeling influence,but there are still some problems such as unclear definition,backward technology and lacking of application.This paper explicitly puts forwardthe research model of social network user influence,combines traditional technology and advanced technology,analyzes the related literature in this field,and mainly discusses the research methods of users influence based on the social network from the perspective of users,content features and deep learning technology,and then further divides it into nature and the neighborhood attri-butes,sentiment analysis and metadata,local network orientedand user and content based characteristics.It also introduces the methods of identification,providing effective and comprehensive reference for scholars in the field.Secondly,it also introduces the data set,evaluation index and experimental results of the prediction application of user influence modeling methods,aiming to predictthe next activation node.Finally,its future development trend is prospected.

Social E-commerce Text Classification Algorithm Based on BERT

LI Ke-yue, CHEN Yi, NIU Shao-zhang

Computer Science. 2021, 48 (2): 87-92. doi:10.11896/jsjkx.200700111

Abstract

PDF(1610KB) ( 2131 )

References | Related Articles | Metrics

With the rapid development of online shopping,a large amount of transaction data has been generated in online transaction activities between online merchants and shoppers,which contain great analytical value.Aiming at the text classification pro-blem of social e-commerce product texts,in order to more efficiently and accurately determine the category of products described in the text,this paper proposes a social e-commerce text classification algorithm based on BERT model.The algorithm adopts the BERT pre-trained language model to complete the feature vector representation of social e-commerce text on sentence-level,and then inputs the obtained feature vectors into the targeted classifier for classification.In this paper,we use the social e-commerce text data set for algorithm verification,and the results show that the F1 value of the trained model on the test set can reach up to 94.61%,which is 6% higher than the MRPC classification task based on the BERT model.Therefore,the social e-commerce text classification algorithm proposed in this paper can more efficiently and accurately determine the type of goods described in the text,which is helpful for further analysis of online transaction data and extraction of valuable information from massive data.

Outlier Detection and Semantic Disambiguation of JSON Document for NoSQL Database

LIU Li-cheng, XU Yi-fan, XIE Gui-cai, DUAN Lei

Computer Science. 2021, 48 (2): 93-99. doi:10.11896/jsjkx.200900039

Abstract

PDF(2977KB) ( 690 )

References | Related Articles | Metrics

With the development of information technology,traditional relational database cannot be used for storage due to multi-source heterogeneity,strong scalability and explosive growth of data in materials and other related fields.Therefore,NoSQL can be used with the charactersitics of schemaless storage and high scalability to solve this problem.As a common data storage format for NoSQL databases,JSON is popular for its simplicity and flexibility.However,NoSQL databases lack schema information,and JSON documents need to be validated and analyzed before being stored in the database.At present,most methods verify the normalization of JSON document format based on JSON schema,which cannot effectively solve the problem of exception detection and semantic ambiguity of JSON document.Therefore,a JSON document outlier detection and semantic disambiguating model for NoSQL database is proposed,named doctorJSON.Based on JSON schema,the model designs outlier detection algorithm deout JSON and semantic disambiguation algorithm disemaJSON to detect the outlier and disambiguation in JSON documents.The vali-dity and efficiency of the model are verified by experiments on the real and synthetic datasets.

Directed Network Representation Method Based on Hierarchical Structure Information

LI Xin-chao, LI Pei-feng, ZHU Qiao-ming

Computer Science. 2021, 48 (2): 100-104. doi:10.11896/jsjkx.191200033

Abstract

PDF(1525KB) ( 664 )

References | Related Articles | Metrics

Network embedding aims to embed each vertex into a low dimensional vector space and preserves certain structural relationships among the vertices in the networks.However,in the directed networks,vertexes can be reached by each other if they are in the same circle,which damages asymmetric transitivity preservation and makes representation learning model hard to capture global information of complex directed networks.This paper proposes an improved representation learning model for directed networks,which weakens the influence of circles in representation learning and enhances the ability of model to obtain global structure information.The proposed method uses TrueSkill to inference hierarchy of a directed graph and compute weight of each edge using hierarchy information.At last,this paper applies this method to some existing embedding models,and then conducts experiments on tasks of link prediction and node classification on several open source datasets.Experimental results show that the proposed method is highly scalable and effective.

k-modes Clustering Guaranteeing Local Differential Privacy

PENG Chun-chun, CHEN Yan-li, XUN Yan-mei

Computer Science. 2021, 48 (2): 105-113. doi:10.11896/jsjkx.200700172

Abstract

PDF(2623KB) ( 844 )

References | Related Articles | Metrics

How to conduct usability data mining while protecting data privacy has become a hot issue.In many practical scena-rios,it is difficult to find a trusted third party to process the sensitive data.This paper proposes the first locally differentially private k-modes mechanism(LDPK-modes) under this distributed scenario.Differing from standard differentially private clustering mechanisms,the proposed mechanism doesn't need any trusted third party to collect and preprocess users data.Users disturb their data using a random response mechanism that satisfies the definition of local d-privacy (local differential privacy with distance metric).When the third party collects the user's disturbed data,it restores its statistical features and generates a synthetic data set.The frequent attributes on the data set are assigned to the initial cluster center and then start k-modes clustering.Theoretical analysis shows that the proposed algorithm satisfies local d-privacy.Experimental results show that our proposal can well preserve the quality of clustering results without a trusted third-party data collector.

User Cold Start Recommendation Model Integrating User Attributes and Item Popularity

HAN Li-feng, CHEN Li

Computer Science. 2021, 48 (2): 114-120. doi:10.11896/jsjkx.200900152

Abstract

PDF(2095KB) ( 885 )

References | Related Articles | Metrics

Cold start has always been a closely watched issue in the field of recommendation systems.Aiming at the problem of cold start for newly registered users,this paper proposes a recommendation model that integrates user demographic information and item popularity.The training set users are divided into several categories by clustering the training set users,and then the distance between the new user and other users in the category is calculated,and the neighboring user set is selected.When calcula-ting the score,we consider comprehensively the impact of popularity,and then push the programs of interest to target users.Finally,the proposed model is verified on the classic recommendation system data set.The results show that the model is significantly better than the traditional collaborative filtering algorithm and has a certain mitigation effect on the cold start problem.

Unsupervised Anomaly Detection Method for High-dimensional Big Data Analysis

ZOU Cheng-ming, CHEN De

Computer Science. 2021, 48 (2): 121-127. doi:10.11896/jsjkx.191100141

Abstract

PDF(2369KB) ( 804 )

References | Related Articles | Metrics

Unsupervised anomaly detection on high-dimensional data is one of the most significant challenges in machine learning.Although previous approaches based on single deep auto-encoder and density estimations have made significant progress,they generate low-dimensional representations as they use only a single deep auto-encoder,indicating that there is insufficient information to perform the subsequent density estimation task.To address the above challenge,a mixed auto-encoding gaussian mixture model (MAGMM) is proposed in this paper.MAGMM substitutes a single deep auto-encoder with a mixture of auto-encoders to generate concatenated low-dimensional representations,so that it can preserve key information from a specific cluster of the input sample.In addition,it utilizes an allocation network to constrain the mixture of auto-encoders,so that each sample can be assigned to a dominant auto-encoder.With the above mechanisms,MAGMM avoids from trapping into local optima and reduces the recons-truction errors,which can facilitate completing the density estimation tasks and improve the accuracy of high-dimensional data anomaly detection.Experimental results show that the proposed method performs better than DAGMM and achieves up to 29% improvement based on the standard F1 score.

Power Load Data Completion Based on Sparse Representation

LI Pei-guan, YU Zhi-yong, HUANG Fang-wan

Computer Science. 2021, 48 (2): 128-133. doi:10.11896/jsjkx.191200152

Abstract

PDF(2303KB) ( 701 )

References | Related Articles | Metrics

Data loss often occurs in the process of power load data collection,which adversely affects the accuracy of algorithm prediction.The existing missing data completion algorithm is only suitable for the case with less missing data,but performs poorly for the case with more missing data.Faced with the challenge of severe data loss,a method for power load missing data completion based on sparse representation is proposed.First of all,we assume that the data is randomly missing,and stitch the assumed missing data in the training data and the complete training data to form a training matrix.Secondly,an over-complete dictionary is generated by discrete cosine transform (DCT),and is learned according to the training matrix,aims to obtain a suitable dictionary for the best sparse representations of the samples in the training matrix.Finally,in the test phase,the upper part of the learned dictionary is used to obtain sparse representations of the missing data in the test set,and then the sparse representations and the lower part of the learned dictionary are used to reconstruct the complete data without missing.Experimental results show that using this method to complete missing values of power load data can achieve higher accuracy than traditional interpolation me-thods,correlation-based KNN algorithm,spatiotemporal compressed sensing estimation algorithm and time-series compressed sen-sing prediction algorithm.Even if the data miss rate is as high as 95%,this method can still effectively complete the missing data.

Image Synthesis with Semantic Region Style Constraint

HU Yu-jie, CHANG Jian-hui, ZHANG Jian

Computer Science. 2021, 48 (2): 134-141. doi:10.11896/jsjkx.200800201

Abstract

PDF(4158KB) ( 951 )

References | Related Articles | Metrics

In recent years,generative adversarial networks have developed rapidly,and image synthesis has become an active research direction.Especially,the combination of semantic region segmentation and generative models provides a new insight for image synthesis.Semantic information can be used to edit and control the input semantic segmentation mask to generate the ideal image with a specific style to generate the desired realistic image.However,the current technology cannot achieve the precise control of the style content of each semantic area.This paper proposes a novel framework for image synthesis under semantic region style constraint,and realizes the adaptive style control of per region using conditional generation model.First of all,a style encoder is used to extract the style information of different semantic regions from the semantic segmentation mask obtained.Then at the generation end,the style information and semantic mask are affine transformed into two sets of modulation parameters respectively for each residual block by using adaptive normalization.The semantic feature map input into the generator is weighted sum according to the modulation parameters,which can effectively combine the semantic information and style information,and gene-rate the target style content gradually through convolution and up-sampling.In the end,this paper designs a new style constraint loss function to constrain the change between per-region style at the semantic level,and to reduce the mutual influence between different semantic style code,aiming at the problem that the existing model cannot accurately control the style of each semantic area.In addition,this paper adopts the method of quantifying weights to compress the generator by about 15.6%,effectively reducing the storage size of the model and the network space without performance degradation.The experimental results show that the proposed model has significantly improved both perceptually and quantitively compared to existing methods,where the FID score is about 3.8% higher than the state-of-the-arts model.

Street Scene Change Detection Based on Multiple Difference Features Network

ZHAN Rui, LEI Yin-jie, CHEN Xun-min, YE Shu-han

Computer Science. 2021, 48 (2): 142-147. doi:10.11896/jsjkx.200500158

Abstract

PDF(3342KB) ( 1085 )

References | Related Articles | Metrics

Street scene change detection plays an important role in the study of natural disaster damage and urban development.Its main goal is to mark out the changing areas in the pair of input images,which is essentially a semantic segmentation problem of binary classification.There may be many interference factors such as light,weather,background noise,viewpoints error and so on when taking street view pictures at different times,which challenges traditional change detection methods.To solve this problem,a new neural network model (Multiple Difference Features Network,MDFNet) is proposed.First,siamese networks are used to extract the different depth features of pairs of input images,and the difference modules are used to calculate the difference of the same depth features to effectively obtain the change information of different depth.Then,by using JPU module to fuse multiple difference features,the deep semantic information can be extracted without losing detail information.Finally,the pyramid pooling module is used to generate the change detection image of the binary classification combined with the global and local information.MDFNet has obtained 0.787 and 0.862 F-scores in the GSV and TSUNAMI part on PCD dataset with 5 fold cross-validation,which are 11.9% and 2.9% higher than the second ranked DOF-CDNet,and can segment the change details more accurately.Therefore,the proposed model can effectively deal with interferences and has an excellent detection ability for complex scenes.

Lightweight Image Retrieval System Based on Feature Clustering

WANG Xiao-fei, ZHOU Chao, LIU Li-gang

Computer Science. 2021, 48 (2): 148-152. doi:10.11896/jsjkx.191200104

Abstract

PDF(2045KB) ( 1065 )

References | Related Articles | Metrics

In the scene of image search,due to the randomness of search request,in order to increase the search speed,it is often necessary to preload the entire data set into the running memory.Because the price of running memory with the same capacity is much higher than that of hard disk,reducing the running memory can greatly reduce the cost of image search service.However,if the data is compressed directly,the search accuracy will be greatly reduced.In this case,this paper proposes a content-based ima-ge search framework,which divides data set into groups.Firstly,the neural network is used to extract image features.On the premise of not compressing the features,a heuristic clustering method is used to group the data,ensuring that there is a certain similarity between the data of each data group.For each data group,HNSW algorithm based on graph structure is used to construct index to speed up image query.In this framework,by controlling the number of data blocks accessed during query,the running memory capacity required by the algorithm can be greatly reduced,under the premise of ensuring the accuracy.

Subset Ratio Dynamic Selection for Consistency Enhancement Evaluation

WANG Kai-xun, LIU Hao, SHEN Gang, SHI Ting-ting

Computer Science. 2021, 48 (2): 153-159. doi:10.11896/jsjkx.200800188

Abstract

PDF(2500KB) ( 529 )

References | Related Articles | Metrics

Due to poor imaging conditions,a lot of underwater images require the consistency enhancement.In the subset-guided consistency enhancement evaluation criterion,the existing subset selection methods need too much subset samples of a whole imageset without any adaptation on data content.Therefore,this paper proposes a subset ratio dynamic selection method for consistency enhancement evaluation.The proposed method further divides the candidate samples into several sampling subsets.Based on a non-replacement sampling strategy,the consistency enhancement degree of an enhancement algorithm is obtained for each sampling subset.By using the student-t distribution under a certain confidence level,the proposed method can adaptively determine the subset ratio for a whole imageset,and the candidate subset is used to predict the consistency enhancement degree of the enhancement algorithm on the whole imageset.Experimental results show that as compared with the existing subset selection me-thods,the proposed method can reduce the subset ratio in all cases,and correctly judge the consistency performance of each enhancement algorithm.With similar evaluation error,the subset ratio of the proposed method can be decreased by 2%~14% over that of the fixed ratio method,and be decreased by 3%~9% over that of the gradual addition method,and thus the complexity is robustly reduced during subset-guided consistency enhancement evaluation.

Multimodal Medical Image Fusion Based on Dual Residual Hyper Densely Networks

WANG Li-fang, WANG Rui-fang, LIN Su-zhen, QIN Pin-le, GAO Yuan, ZHANG Jin

Computer Science. 2021, 48 (2): 160-166. doi:10.11896/jsjkx.200400095

Abstract

PDF(3238KB) ( 1060 )

References | Related Articles | Metrics

The image fusion method based on residual network and dense network has the problem of losing some useful information in the middle layer of network and unclear details of fusion image.Therefore,a multi-modal medical image fusion based on the Dual Residual Hyper-Densely Networks (DRHDNs) is proposed.DRHDNs is divided into two parts:feature extraction and feature fusion.In the feature extraction part,a dual residual hyper dense blocks is constructed by combining hyper dense connection and residual learning.The hyper dense connection not only occurs between layers in the same path,but also between layers in different paths.This connection makes the feature more sufficient,the detail information more abundant,and the initial feature fusion of the source image is carried out .Feature fusion part is for final fusion.Compared with the other six image fusion methods,four groups of brain images are fused and compared,and an objective comparison is made from four evaluation indexes.Results show that DRHDNs has good performance in detail retention and contrast.The fusion image has rich and clear detail information,which conforms to human visual.

Event-based User Experience Evaluation Method for Virtual Reality Applications

MA Si-qi, CHE Xiao-ping, YU Qi, YUE Chen-feng

Computer Science. 2021, 48 (2): 167-174. doi:10.11896/jsjkx.200100065

Abstract

PDF(2596KB) ( 1188 )

References | Related Articles | Metrics

With the rapid development of virtual reality(VR) technology in recent years,the VR industry is currently growing fast.Although VR provides many new possibilities for content form,due to the immaturity of VR application design,designers in the VR industry still face severe challenges.Design of VR applications still needs to be verified by users.Therefore,user expe-rience(UX) analysis is really critical to the success of VR software development.At the same time,with the increase of users' requirements for VR content quality,whether the content is attractive enough will greatly affects users' experience and determine the service life of VR applications.Consequently,searching for interaction events that will affect VR UX is critical to improve user stickiness.At present,the researches on VR mainly focus on the improvement of hardware and software,and pay less attention to the content event design.Moreover,there is no unified and clear standard for VR UX evaluation.This paper attempts to find out the relationship between user traits,VR interaction events and user experience through realistic experiments.This paper firstly defines four types of VR interaction events,and designs a questionnaire for collecting the tester's traits and their subjective evaluation.During the experiment,objective physiological data of the testers and their participation process are recorded.The 80 testers are divided into two groups to experience two types of VR games with constant time.The statistical method and improved Prism algorithm are used to find out the correlation among user traits,type of game interaction events and user experience.The experiment results can provide references for VR designers and developers,and at the same time provide the preliminary study to standardized VR user experience evaluation.

Knowledge Graph Construction Techniques:Taxonomy,Survey and Future Directions

HANG Ting-ting, FENG Jun, LU Jia-min

Computer Science. 2021, 48 (2): 175-189. doi:10.11896/jsjkx.200700010

Abstract

PDF(2659KB) ( 5225 )

References | Related Articles | Metrics

With the concept of knowledge graph proposed by Google in 2012,it has gradually become a research hotspot in the field of artificial intelligence and played a role in applications such as information retrieval,question answering,and decision analysis.While the knowledge graph shows its potential in various fields,it is easy to find that there is no mature knowledge graph construction platform currently.Therefore,it is essential to research the knowledge graph construction system to meet the application needs of different industries.This paper focuses on the construction of the knowledge graph.Firstly,it introduces the current mainstream general knowledge graphs and domain knowledge graphs and describes the differences between the two in the construction process.Then,it discusses the problems and challenges in the construction of the knowledge graph according to various types.To address the above-mentioned issues and challenges,it describes the five-level solution methods and strategies of knowledge extraction,knowledge representation,knowledge fusion,knowledge reasoning,and knowledge storage in the current graph construction process.Finally,it discusses the possible directions for future research on the knowledge graph and its application.

Development of Lévy Flight and Its Application in Intelligent Optimization Algorithm

ZHENG Jie-feng, ZHAN Hong-wu, HUANG Wei, ZHANG Heng, WU Zhou-xin

Computer Science. 2021, 48 (2): 190-206. doi:10.11896/jsjkx.200500142

Abstract

PDF(2199KB) ( 976 )

References | Related Articles | Metrics

Lévy Flight,originated from pure mathematical concepts,has been widely used in many fields,such as physics,biology,statistics,finance and computer science.At present,there is no summary of its development and application in intelligent optimization algorithm in China.Therefore,this paper reviews the development and application of Lévy Flight,and introduces the basic principle and application of Lévy Flight related variants.Then it focuses on the study of applying Lévy Flight to intelligent optimization algorithm in recent ten years,and classifies and analyzes its application methods.Finally,the future development trend of Lévy Flight is summarized.The purpose of the review is to let researchers understand the basic principle of Lévy Flight and its development in intelligent optimization algorithm,and to promote the development and application of Lévy Flight and its variants in many disciplines,especially in computer science.

Evaluation of Quality of Interaction in Online Learning Based on Representation Learning

WANG Xue-cen, ZHANG Yu, LIU Ying-jie, YU Ge

Computer Science. 2021, 48 (2): 207-211. doi:10.11896/jsjkx.201000042

Abstract

PDF(1735KB) ( 826 )

References | Related Articles | Metrics

The model of education today has undergone a very significant change,and education is developing in the direction of ubiquity,intelligence and individuation.Online education,represented by MOOCs,is gradually coming into the public field of vision,and the interactivity in online education has become the key to determine the quality of online learning.Some researches show that the interaction in the learning process provides efficient help and effective support for learners,and the feedback of learning process evaluation can effectively improve the interaction effect of learning.Modeling interactions between learners and learning resources is crucial in domains such as e-commerce.Representation learning presents a method to model the sequential interactions between learners and learning resources.Firstly,an interactive network of online learning is established.And then,the learners and learning resources can be embedded into a Euclidean space by using two recurrent neural networks.The evaluation index of the quality of interaction is proposed,which can judge whether the learner's learning effect is up to the expectation.The experiments on real datasets reveal the effectiveness of the proposed method.

Dialogue Act Prediction Based on Response Generation

WANG Bo-yu, WANG Zhong-qing, ZHOU Guo-dong

Computer Science. 2021, 48 (2): 212-216. doi:10.11896/jsjkx.200700137

Abstract

PDF(1875KB) ( 1043 )

References | Related Articles | Metrics

With the continuous development of the human-machine dialogue system,it is of great significance for the computer to accurately understand the speaker's dialogue act and predict the act of response according to the history information of the dialogue.Previous research work focus on act prediction of responses based on dialogue text and existing labels.But in many scena-rios,the reply has not been generated.Therefore,this paper proposes a dialogue act prediction model based on reply generation.In the generation part,the Seq2Seq structure is used to generate text based on the conversation history information as text information for future replies in the conversation;in the classification part,the LSTM model is used to express the generated reply text and the existing conversation information as clause level representations.Combined with the attention mechanism,it highlights the connection between the dialogue sentence of the same round and the generated response.The experimental results show that the proposed model a chieves a 2.54% F1-score improvement compared to the simple baseline model,and the joint training method contributes to the improvement of model performance.

Unsupervised Domain Adaptation Based on Weighting Dual Biases

MA Chuang, TIAN Qing, SUN He-yang, CAO Meng, MA Ting-huai

Computer Science. 2021, 48 (2): 217-223. doi:10.11896/jsjkx.200700028

Abstract

PDF(1759KB) ( 650 )

References | Related Articles | Metrics

Unsupervised domain adaptation (UDA) is a new kind of machine learning paradigm,which facilitates the training of target domain model through transferring knowledge from source domain to unlabeled target domain.In order to model the domain distribution difference between the source domain and target domain,the maximum mean discrepancy (MMD) is widely applied,it plays an effective role in promoting the performance of UDA.Usually,the class size and data distribution between the target domain and the source domain are not the same,unfortunately,these methods usually ignore this structure information.To this end,this paper proposes a model called sample weighted and class weighted based unsupervised domain adaptation network (SCUDAN).On one hand,the class distribution alignment between the source domain and the target domain is achieved through adaptive weighting on the classes of the source domain.On the other hand,the class centers between the target domain and the source domain can be aligned through adaptive weighting on the samples of the target domain.In addition,a CEM (Classification Expectation Maximization) algorithm is proposed to optimize SCUDAN.Finally,the effectiveness of the proposed method is verified by comparative experiments and analysis.

Method for Prediction of Red Blood Cells Supply Based on Improved Grasshopper Optimization Algorithm

LIU Qi, CHEN Hong-mei, LUO Chuan

Computer Science. 2021, 48 (2): 224-230. doi:10.11896/jsjkx.200600016

Abstract

PDF(1511KB) ( 532 )

References | Related Articles | Metrics

At present,the problem of blood supply shortage is quite serious.There exists the phenomenon that short supply happen between blood stations and institutions that use blood.Aiming at such a problem,the LSTM prediction method based on the improved grasshopper optimization algorithm(GOA) is proposed in order to predict red blood cells supply in the future and provide effective guidance for workers in making blood collection plan and preparation plan.By using LSTM to capture the under-lying patterns between the historical data,the effect of predicting the future can be achieved.There are two parts of work.Firstly,aiming at the problem that the conventional GOA is easy to fall into local optimum and has a slower convergence speed,the model of refracting opposite-based learning and chaotic mapping are introduced to GOA so as to improve the global exploration capability.Secondly,in order to improve the accuracy of LSTM,it is combined with the improved GOA and evaluate the perfor-mance of the improved LSTM model by using the real data of red blood cells supply in a certain area.Comparing to the conventional LSTM,the MAE,MAPE,RMSE are reduced by 39.827 8,1.10%,55.819 1,respectively.The experimental results show that the proposed method has higher reliability.

Deaf Sign Language Recognition Based on Inertial Sensor Fusion Control Algorithm

RAN Meng-yuan, LIU Li, LI Yan-de, WANG Shan-shan

Computer Science. 2021, 48 (2): 231-237. doi:10.11896/jsjkx.191200143

Abstract

PDF(3176KB) ( 801 )

References | Related Articles | Metrics

How to effectively communicate with the outside world for deaf people has always been a difficult issue that has attracted much attention.This paper proposes a sign language recognition scheme based on inertial sensor fusion control algorithm,which aims to achieve efficient and accurate real-time sign language recognition.This fusion control algorithm uses feedback control ideas to fuse two traditional attitude information calculation methods,reduces the impact of the environment on sensors,and can accurately obtain the attitude information of the measured object in the transient state.The algorithm performs data fusion,data preprocessing and feature extraction on the collected deaf-mute sign language data,and uses an adaptive model integration method composed of support vector machines(SVM),K-nearest neighbors(KNN) and feedforward neural networks(FNN) for classification.The results show that the proposed sensor fusion control algorithm effectively obtains real-time poses.The sign language recognition scheme achieves accurate recognition of 30 deaf-mute pinyin sign languages with a recognition accuracy of 96.5%.The work of this paper will lay a solid foundation for sign language recognition of deaf and dumb people,and provide refe-rences for related research on sensor fusion control.

Data Efficient Third-person Imitation Learning Method

JIANG Chong, ZHANG Zong-zhang, CHEN Zi-xuan, ZHU Jia-cheng, JIANG Jun-peng

Computer Science. 2021, 48 (2): 238-244. doi:10.11896/jsjkx.191100107

Abstract

PDF(1886KB) ( 740 )

References | Related Articles | Metrics

Imitation learning provides a framework to make agent learn an efficient policy from expert demonstrations.During the learning process,the agent does not need to interact with the expert or get access to an explicit reward signal,but only needs a large number of expert demonstrations.Classical imitation learning methods usually need to imitate from first-person expert demonstrations,a sequence of states and actions that expert should have taken.However,most expert demonstrations exist in the form of third-person videos in reality.Different from the first-person expert demonstrations,there is a difference between the viewpoint of the third-person demonstrations and samples generated by the agent,resulting in a lack of one-to-one correspondence between them.Therefore,the third-person demonstrations cannot be directly used in imitation learning.To alleviate this problem,this paper presents a data efficient third-person imitation learning method.Firstly,this method introduces the image difference based on Generative Adversarial Imitation Learning(GAIL) to eliminate the domain features including the background of environment and colors by taking advantage of the Markov property of Markov decision process and the time continuity of states.And the most relevant part of policy can be achieved for imitation learning.Secondly,this paper introduces a variational discriminator bottleneck to limit the discriminator to alleviate the influence of domain features on the process of learning policy.In order to verify the performance of the proposed algorithm,this paper makes experiments on three MuJoCo tasks,and compares it with the existing algorithms.Experimental results indicate that the proposed method can achieve significant performance improvements over existing methods and does not require additional demonstrations,when dealing with imitation learning from third-person expert demonstrations.

Recurrent Convolution Attention Model for Sentiment Classification

CHEN Qian, CHE Miao-miao, GUO Xin, WANG Su-ge

Computer Science. 2021, 48 (2): 245-249. doi:10.11896/jsjkx.200100078

Abstract

PDF(1980KB) ( 872 )

References | Related Articles | Metrics

Sentiment classification has important application value for downstream applications,including recommendation system,automatic question answering and reading comprehension.It is an important research direction in the field of natural language processing.The task of sentiment classification depends on global and local information hidden in context.However,exis-ting neural network models can not capture the local and global information of context at the same time.In this paper,a recurrent convolutional attention model (LSTM-CNN-ATT,LCA) is proposed for single label and multi-label sentiment classification tasks.It uses attention mechanism to fuse the local information extraction ability of convolutional neural network and the global information extraction ability of recurrent neural network,including word embedding layer,context representation layer,convolution layer and attention layer.For the multi-label sentiment classification task,the topic information is added to the attention layer to further guide the accurate extraction of multi-label emotion tendency.The F1 index on two single label datasets reaches 82.1%,which is equivalent to the frontier single label model.On two multi-label datasets,the experimental results on small datasets are close to the benchmark model,and the F1 index on large datasets reaches 78.38%,which is higher than the state-of-the-art model.It indicates that LCA model has high stability and strong universality.

Artificial Potential Field Path Planning Algorithm for Unknown Environment and Dynamic Obstacles

DU Wan-ru, WANG Xiao-yin, TIAN Tao, ZHANG Yue

Computer Science. 2021, 48 (2): 250-256. doi:10.11896/jsjkx.191100170

Abstract

PDF(3189KB) ( 1488 )

References | Related Articles | Metrics

The actual battlefield environment is complex.Many hidden and dynamic obstacles cannot be detected in advance by means of high altitude.It is a threat to the security of the agent.Aiming at the unknown battlefield environment with various obstacles,taking avoiding static and dynamic obstacles and tracking targets as the research object,an APF(Artificial Potential Field) path planning algorithm for unknown environment and dynamic obstacles is proposed.In this algorithm,the agent constructs the gravitational potential field centered on the target point and the repulsive potential field centered on the obstacle,perceives the motion information of the local obstacle and the target point on the route of the agent,and adds the information into the calculation of the potential field function to achieve the effect of dynamic obstacle avoidance and tracking.On the other hand,it introduces the distance factor and dynamic temporary target point to eliminate the minimum solution and path jitter of APF algorithm.The results show that the proposed algorithm can avoid dynamic obstacles and track the target points flexibly in unknown environment,and can effectively eliminate the dead solution and path jitter problems.The proposed algorithm is compared with the traditional APF algorithm and the algorithm described in literature [19] with a dynamic obstacle avoidance mechanism added.Experimental results show that the APF algorithm can successfully resolve the problem of path planning failure of the two comparative algorithms and successfully complete the task of path planning,and the success rate is more than 95%.

Simulation Analysis on Dynamic Ridesharing Efficiency of Shared Autonomous Taxi

ZENG Wei-liang, HAN Yu, HE Jin-yuan, WU Miao-sen, SUN Wei-jun

Computer Science. 2021, 48 (2): 257-263. doi:10.11896/jsjkx.200400008

Abstract

PDF(3209KB) ( 656 )

References | Related Articles | Metrics

Shared autonomous taxi is one of the revolutionary intelligent transportation modes in the future,which will produce huge social and environmental benefits.The maximum number of rideshare is a key parameter affecting passengers' travel time,price,comfort and operating cost.However,previous researches rarely analyzed the maximum number of rideshare.To fill this gap,a dynamic autonomous taxi simulation system is developed.It consists of three models:searching,scheduling and waiting,and investigates how the maximum number of rideshare influences the system performance under the changing travel demand.The road network of the Nanshan district in Shenzhenis examined as a case study to evaluate the ridesharing efficiency in different settings of the maximum number of rideshare and the travel demand.The simulation results show that switching from traditional taxis to shared autonomous taxis can greatly increase the success rate of the serviced requests by 20% and reduce the total travel time by 3％～23%.Interestingly,the ridesharing efficiency converges gradually as the maximum number of rideshare increasing to a certain value.The ridesharing efficiency can be almost optimized when the maximum number of rideshare is set to 3 or 4 for the case of high travel demand.It can be concluded that multi passenger ridesharing can alleviate the current problem of struggle to hail a taxi,and as the travel demand increases,the shared autonomous taxis system has a stronger robustness compared with traditional non-shared taxi system.

Traffic Flow Forecasting Method Combining Spatio-Temporal Correlations and Social Events

LYU Ming-qi, HONG Zhao-xiong, CHEN Tie-ming

Computer Science. 2021, 48 (2): 264-270. doi:10.11896/jsjkx.200300098

Abstract

PDF(1721KB) ( 1045 )

References | Related Articles | Metrics

Traffic flow prediction,as a key issue in intelligent transportation system,becomes a research hotspot in the field of transportation both at home and abroad.The main challenge of traffic flow prediction is twofold.First,traffic flow has complica-ted spatial and temporal correlations.Second,traffic flow can be influenced by social events.Aiming at these challenges,this paper proposes a deep learning framework for traffic flow prediction.On the one hand,a sub-network by combining graph convolutional neural network and recurrent neural network is designed to extract spatio-temporal correlation features from the non-European road network space.On the other hand,a sub-network based on convolutional neural network is designed to extract social event features from textual data.Finally,the traffic flow prediction model is implemented by merging the spatio-temporal correlation feature extraction subnetwork and the social event feature extraction sub-network.In order to verify the validity of the model,experiments are conducted based on real traffic flow data.Compared with the baseline methods,the proposed method has higher accuracy,and the accuracy improves by 3% to 6%.

Summary of Principle and Application of Blockchain

GUO Shang-tong, WANG Rui-jin, ZHANG Feng-li

Computer Science. 2021, 48 (2): 271-281. doi:10.11896/jsjkx.200800021

Abstract

PDF(1938KB) ( 3467 )

References | Related Articles | Metrics

In recent years,as digital cryptocurrency has gradually come into people's sight,its underlying technology blockchain has also attracted people's attention.As a distributed ledger technology,blockchain is characterized by multi-party maintenance,non-tampering,openness and transparency.In this paper,the structure of block chain is divided according to the hierarchy,and the functions and principles of each layer are introduced from low to high.Block chain is divided into public chain,alliance chain and private chain according to the degree of openness.The working principle of public chain and alliance chain is illustrated by taking Bitcoin and Hyperledger Fabric as examples.And this paper gives a detailed introduction to the underlying core technology consensus algorithm,smart contract and privacy security of blockchain,and analyzes the research progress and research prospect of blockchain in the end.

Energy Classifier Based Cooperative Spectrum Sensing Algorithm for Anti-SSDF Attack

DING Shi-ming, WANG Tian-jing, SHEN Hang, BAI Guang-wei

Computer Science. 2021, 48 (2): 282-288. doi:10.11896/jsjkx.191100124

Abstract

PDF(2711KB) ( 500 )

References | Related Articles | Metrics

Spectrum sensing is an important part of cognitive radio communication and is under the serious threats of spectrum sensing data falsification (SSDF) in terms of security,which trends dynamically in the attacking methods nowadays.It is difficult to identify dynamic tampered data using traditional defense algorithms because the traditional ways always assume that the atta-cking strength remains unchanged.Aiming at the dynamic SSDF attack,an energy classifier enabled cooperative spectrum sensing algorithm for anti-SSDF attack is proposed.This algorithm firstly analyzes the characteristics of dynamic SSDF attacks,combines with the distance discriminant method to classify the neighbor users.Then it identifies the malicious neighbor users by comparing the classification results with the local results.Afterward the local user establishes a reputation model under the sliding time window based on the information of historical results,thereby updates the reputation values of neighbor users and will eventually implement weighted cooperative spectrum sensing.The simulation results show that compared with the largest deviation-based distributed cooperative spectrum sensing (LDCSS) algorithm and the reputation-based cooperative spectrum sensing (RBCSS) algorithm,this algorithm provides an increased spectrum detection probability by 15% and 16% respectively when the attacking strength is close to the threshold,which not only significantly increases the cooperative spectrum sensing performance of the cognitive network,but also improves the efficiency of spectrum sharing.

High-speed Replay of Ethereum Smart Contracts Based on Block Parallel

CHEN Zi-min, LU Yi-wen, GUO Yan

Computer Science. 2021, 48 (2): 289-294. doi:10.11896/jsjkx.200500105

Abstract

PDF(2609KB) ( 2036 )

References | Related Articles | Metrics

Analyzing and researching blocks,transactions,accounts,and smart contract data on Ethereum is of great value,but Ethereum has a large amount of data,many types of data,and different storage structures.The current data acquisition methods are slow and the acquired data are incomplete,so it is very difficult to use these data.This paper proposes Geth-query,a fast data export tool for Ethereum based on block parallel.By analyzing the internal mechanism of Ethereum,it uses snapshots of the world state of the block to eliminate dependencies between blocks and optimize the efficiency of local resource utilization to parallel replay block,thus achieving fast and comprehensive extraction of data on the Ethereum chain.Experiments prove that the types of data extracted by Geth-query are rich,and the data export speed is about 10 times faster than traditional methods.For ease of use,this paper also optimizes the storage of the exported data and displays the data on the front-end page,thus providing a data foundation for the analysis and research of Ethereum.

Intelligent Manufacturing Security Model Based on Improved Blockchain

WANG Wei-hong, CHEN Zhen-yu

Computer Science. 2021, 48 (2): 295-302. doi:10.11896/jsjkx.191200159

Abstract

PDF(3570KB) ( 787 )

References | Related Articles | Metrics

In view of the traditional block chain intelligent manufacturing security model's slow speed of block construction and data query,and high time complexity of inserting query operation,an intelligent manufacturing security model based on improved block chain is proposed.Firstly,the disadvantages of traditional block chain are solved,such as large power consumption and low throughput.A new Merkle Patricia tree (MPT) is introduced to expand the block chain structure to provide fast query of node state.Aiming at the problem that MPT does not support concurrent operation and poor performance under high load state,Merkle is designed as a lockless concurrent cache Patricia tree,which supports concurrent data operation without lock,and can improve the efficiency in multi-core system.Finally,the performance of the proposed model is analyzed by specific simulation experiments.The results show that the improved intelligent manufacturing security model of block chain can effectively reduce the time complexity of insertion query operation,greatly improve the speed of block construction and data query,and compared with the traditional model,it can get better overall performance.

Improved PBFT Consensus Algorithm Based on Trust Matching

JI Yu-xiang, HUANG Jian-hua, WANG Zhe, ZHENG Hong, TANG Rui-cong

Computer Science. 2021, 48 (2): 303-310. doi:10.11896/jsjkx.200500112

Abstract

PDF(1978KB) ( 1419 )

References | Related Articles | Metrics

Consensus algorithm is the key to realize data consistency in decentralized blockchain systems.Aiming at the scalability and security problems of Practical Byzantine Fault Tolerance (PBFT),a Trust-based Matching Byzantine Fault Tolerance (TMBFT) algorithm is proposed.Firstly,the trust-based neighbor matching model is used to select some nodes for voting consensus,so as to reduce the traffic of the blockchain network.Secondly,a trust evaluation mechanism is introduced to supervise the behavior of neighbor nodes,to ensure the effective detection of Byzantine nodes and the security of node voting.Finally,a vote counting mechanism is designed to ensure the consistency of consensus results and improve the efficiency of consensus.Compared with PBFT,TMBFT reduces the communication complexity from O(N²) to O(Nlog₂N),and effectively reduces the communication overhead in the network.Security analysis shows that the trust evaluation mechanism reduces the probability of malicious voting and improves the system security effectively.Experimental results show that TMBFT has better performance than the traditional Byzantine algorithm.

Model Chain for Data and Model Sharing

YAN Kai-lun, ZHANG Ji-lian

Computer Science. 2021, 48 (2): 311-316. doi:10.11896/jsjkx.191000126

Abstract

PDF(2601KB) ( 823 )

References | Related Articles | Metrics

Machine learning has been applied in more and more scenarios,but software that employs machine learning to perform tasks depends on third-party to update the models.This paper proposes and implements a model chain by utilizing computation power of training neural network consumption with proof-of-work.As a blockchain that can be used to share data and machine learning models,the data shared anonymously by the whole network node are used in the model chain,and the neural network model is explored based on the primary network,thus realizing neural network model update without relying on the third-party.The shared data are signed with a ring signature to protect local data privacy.The whole network uses the same test set to evaluate the model,and the adopted model can be regarded as proof-of-work.This paper proposes two reward mechanisms,i.e.,material reward and model reward.To deal with potential threats,e.g.,blockchain ledger analysis,dirty data attacks and fraudulent voting,this paper proposes ideal ring signature scheme and several solutions.Finally,extensive experiments on real data are conducted,and the results show that the model in the model chain can adapt to the user changes and data changes.

Reconstruction of Cloud Platform Attack Scenario Based on Causal Knowledge and Temporal- Spatial Correlation

WANG Wen-juan, DU Xue-hui, REN Zhi-yu, SHAN Di-bin

Computer Science. 2021, 48 (2): 317-323. doi:10.11896/jsjkx.191200172

Abstract

PDF(2670KB) ( 931 )

References | Related Articles | Metrics

Attack behavior in cloud computing environment gradually shows characteristics of strong concealment and complex multi-step,that is,a complete attack needs to execute some different attack steps to achieve the final goal.However,the existing intrusion detection system usually does not have the necessary ability of correlation,and can only detect single-step attack or attack fragment,so it is difficult to find and identify multi-step attack,and unable to restore attackers' attack process completely.To solve this problem,this paper proposes an attack scenario reconstruction technique based on causal knowledge and space-time correlation.Firstly,the bayesian network is used to model the causal knowledge,and the causal attack patterns are extracted from the alerts with IP address correlation,so as to provide template basis for the subsequent correlation analysis.Then,on the basis of causal knowledge network,alert correlation is conducted from the perspectives of causal,temporal and spatial dimensions to discover potential hidden relationships,and high-level attack scenarios are reconstructed to provide basis and reference for building a cloud environment that can be monitored and accountable.

IoTGuardEye:A Web Attack Detection Method for IoT Services

LIU Xin, HUANG Yuan-yuan, LIU Zi-ang, ZHOU Rui

Computer Science. 2021, 48 (2): 324-329. doi:10.11896/jsjkx.200800030

Abstract

PDF(2689KB) ( 992 )

References | Related Articles | Metrics

In most of the edge computing applications including Internet of Things (IoT) devices,the application programming interface (API) based on Internet application technologies,which are commonly known as Web Technologies,is the core of information interaction between devices and remote servers.Compared with traditional web applications,most users cannot directly access APIs used by edge devices,which makes them suffer fewer attacks.However,with the popularity of edge computing,the attack based on API has gradually become a hot spot.Therefore,this paper proposes a web attack vector detection method for IoT service providers.It can be utilized to detect malicious traffic against its API services and provide security intelligence for the security operation center (SOC).Based on the feature extraction of text sequence requested by hypertext transfer protocol (HTTP),this method combines bidirectional long short-term memory (BLSTM) to detect the attack vector of web traffic according to the relatively fixed format of API request message.Experimental results show that,compared with the rule-based Web application firewall (WAF) and traditional machine learning methods,the proposed method has better recognition ability for attacks on IoT service APIs.

Cross-domain Few-shot Face Spoofing Detection Method Based on Deep Feature Augmentation

SUN Wen-yun, JIN Zhong, ZHAO Hai-tao, CHEN Chang-sheng

Computer Science. 2021, 48 (2): 330-336. doi:10.11896/jsjkx.200100020

Abstract

PDF(2173KB) ( 747 )

References | Related Articles | Metrics

The face recognition technology is improving rapidly these days.One the other side,the face presentation attack has become a practical security problem.To protect the system,face presentation attack detection methods are employed for detecting such attacks in advance.This paper extends a classic domain adaptation method to the deep neural network scenario,defines a feature augmentation-based domain adaptation layer,proposes a cross-domain few-shot face presentation attack detection method based on deep feature augmentation.This method is based on the existing method based on Fully Convolutional Network and improves the existing method by embedding a domain adaptation layer in the middle of the network.The new layer augments the feature maps,adapts the difference between the source and target domains.Then,a pixel-level probability map is predicted based on the augmented the feature maps.Finally,the prediction map is fused to a frame-level decision.Experiments are conducted on the CASIA-FASD,Replay-Attack and OULU-NPU datasets.Six commonly used protocols including the cross-dataset protocols between CASIA-FASD and Replay-Attack,the standard protocols of the OULU-NPU dataset are followed.The training and test data are cross different backgrounds,presentation attack instruments and cameras.The experiment results show that the baseline method,the Fully Convolutional Networkbased face presentation attack detection method has already achieved state-of-the-art performance.The performance can be further improved by learning the domain adaptation model on small-sample data in the target domain.The proposed method can halve the error rate by introducing domain adaptation (train on CASIA-FASD and test on Replay-Attack:decreased from 27.31% to 11.23%,train on Replay-Attack and test on CASIA-FASD:decreased from 37.33% to 21.83%,OULU-NPU's standard protocol IV:decreased from 9.45% to 5.56%).This confirms the effectiveness of the proposed method.