Computer Science

ZHONG Yue, GU Jieming, CAO Honglin

Computer Science. 2023, 50 (9): 3-15. doi:10.11896/jsjkx.230500190

Abstract

PDF(3192KB) ( 2476 )

References | Related Articles | Metrics

With the rapid development of information technology,human beings are entering the era of ubiquitous connectivity,where billions of Internet of Things(IoT) devices are connected to the network.The continuous growth of network attacks targeting user privacy and the network environment has made it crucial to ensure the information security of IoT devices.Due to the limited computational capabilities,battery capacity,and memory resources of IoT devices,conventional block cipher algorithms are not suitable for IoT devices that require low latency and low power consumption,lightweight block cipher algorithms have emerged to address these challenges.This paper provides an overview of the research status and progress of lightweight block cipher algorithms,and categorizes them into six types according to their structure.It comprehensively compares and analyzes the hardware and software implementations of lightweight block cipher algorithms based on multidimensional evaluation criteria.Furthermore,it explores the security,resource consumption,and performance aspects in-depth.Finally,this paper discusses the future research directions of lightweight block cipher algorithms.

Study on Blockchain Based Access Control Model for Cloud Data

TONG Fei, SHAO Ranran

Computer Science. 2023, 50 (9): 16-25. doi:10.11896/jsjkx.230500239

Abstract

PDF(2344KB) ( 2135 )

References | Related Articles | Metrics

The combination of blockchain and ciphertext policy sttribute based rncryption(CP-ABE) schemes has been widely used in the access control of sharing data on the cloud,but the privacy protection of data users in these schemes has not been solved.Some studies introduce distributed multi-authority attribute based signature schemes(DMA-ABS) to protect the privacy of data users,but when the data user accesses the data multiple times,it is necessary to perform repeated permission verification,which will cause unnecessary time consumption.And when the attributes and access control policies of data users are relatively unchanged,data users can access shared data repeatedly and infinitely,system overload and affect normal request processing.This may cause the leakage of cloud data, posing a hidden danger to the security of cloud data.At the same time,the behavior of data users changes dynamically.A data user who once perform well may have some malicious behaviors such as frequent access to data,illegal access to data,which brings hidden dangers to data security.Firstly,the smart contract is combined with the CP-ABE scheme of multi-attribute authority center to realize the fine-grained access control of personal privacy data in the cloud,and the distributed multi-authority attribute based signature scheme is introduced.The anonymous identity verification of data users is completed to protect the identity privacy of data users.Secondly,based on the idea of unspent transaction output(UTXO ) of Bitcoin,the digital token is designed to realize once authorization and multiple access.Finally,this scheme implements an access control process based on hyperledger fabric,and compares it with existing schemes in terms of access time overhead.The results indicate that the proposed scheme can effectively reduce access time overhead and improve access efficiency.

Efficient Encrypted Image Content Retrieval System Based on SecureCNN

LU Yuhan, CHEN Liquan, WANG Yu, HU Zhiyuan

Computer Science. 2023, 50 (9): 26-34. doi:10.11896/jsjkx.230400033

Abstract

PDF(2100KB) ( 2190 )

References | Related Articles | Metrics

With the rapid development of smart devices,content-based image retrieval technology(CBIR) on the cloud is becoming increasingly popular.However,image retrieval on a semi-honest cloud server carries the risk of compromising user privacy.To prevent personal privacy from being compromised,users encrypt their images before outsourcing them to the cloud,but existing CBIR schemes on plaintext domains are ineffective for searching encrypted image data.To solve these problems,an efficient encrypted image content retrieval scheme based on approximate number homomorphism is proposed in the paper,which can quickly achieve image search without continuous user interaction while protecting user privacy.Firstly,feature extraction of image sets using approximate number homomorphism neural network can ensure that the parameters of the network model and the image set data are not leaked to the cloud server.Secondly,a new neural network partitioning method is also proposed to reduce the homomorphic encryption multiplication depth and improve the model operation efficiency,and also construct the index using hierarchical navigable small world(HNSW) algorithm to achieve efficient image retrieval.In addition,homomorphic encryption is used to guarantee the security of image data transmission process and symmetric encryption algorithm is used to guarantee the security of storage stage.Finally,the security and efficiency of the scheme are proved by experimental comparison and security analysis.Experimental results show that the scheme is IND-CCA,and the number of multiplications of homomorphic encryption in this scheme is at most 3 times while guaranteeing the image privacy,which far exceeds the existing schemes in terms of retrieval accuracy and at least 100 times higher than the existing schemes in terms of retrieval time complexity,achieving a balance of retrieval accuracy and efficiency.

Hierarchical Task Network Planning Based Attack Path Discovery

WANG Zibo, ZHANG Yaofang, CHEN Yilu, LIU Hongri, WANG Bailing, WANG Chonghua

Computer Science. 2023, 50 (9): 35-43. doi:10.11896/jsjkx.230500025

Abstract

PDF(3616KB) ( 2047 )

References | Related Articles | Metrics

Attack path discovery is a critical task for cyber asset security assessment.The existing artificial intelligence-based planning for attack path discovery method is favored by security practitioners due to its rich modeling language and complete planning algorithm,but its scalability problem cannot be ignored.For that reason,a hierarchical task network-based attack path method is proposed.Specifically,concerning the scalability problem,the proposed method is decomposed into the following three stages:first and foremost,focusing on undesirable attack path generation performance caused by expanding network scale,a multi-level K-way partitioning algorithm is introduced into the target topology; subsequently,focusing on the difficulty of domain problem description caused by complex discovery tasks,an attack path-oriented hierarchical task network is constructed with the combination of expert experiments; and finally,focusing on low attack path updating efficiency caused by demands on what-if security analysis,a maintenance scheme is designed for local information changes of assets.Experimental results show that the proposed method is suitable for attack path discovery in large-scale network which has an advantage over efficiency and scalability.

Network Protocol Vulnerability Mining Method Based on the Combination of Generative AdversarialNetwork and Mutation Strategy

ZHUANG Yuan, CAO Wenfang, SUN Guokai, SUN Jianguo, SHEN Linshan, YOU Yang, WANG Xiaopeng, ZHANG Yunhai

Computer Science. 2023, 50 (9): 44-51. doi:10.11896/jsjkx.230600013

Abstract

PDF(2757KB) ( 2016 )

References | Related Articles | Metrics

With the deep integration of informatization and industrialization,the security issues of industrial Internet of things(IIoT) network protocols are becoming increasingly prominent.Existing network protocol vulnerability mining techniques mainly relyon feature variation and fuzzy testing,which have the limitations of depending on expert experience and cannot overcome the challenges posed by unknown protocols.To address the vulnerability mining challenges in IIoT protocols,this paper conducts research on the automation analysis and generation of vulnerability detection rules and proposes a network protocol vulnerability mining method based on a combination of generative adversarial networks(GANs) and mutation strategies.Firstly,a network protocol analysis model based on GANs is employed to conduct deep information mining on message sequences,extract message formats,and related features,enabling the recognition of network protocol structures.Then,by combining a guided iterative mutation strategy with a mutation operator library,directed test case generation rules are constructed to reduce the time for vulnerabi-lity discovery.Ultimately,an automated vulnerability mining method for unknown industrial control network protocols is deve-loped to meet the demand for protocol automated vulnerability mining in the existing industrial control application domain.Based on the above-mentioned approach,we conduct tests on two industrial control protocols(Modbus-TCP and S7) and evaluate them in terms of test coverage,vulnerability detection capability,test case generation time,and diversity.Experimental results show that the proposed method achieves a remarkable 89.4％ on the TA index.The AD index,which measures the ability to detect vu-lnerabilities in the simulated ModbusSlave system,reaches 6.87％.Additionally,the proposed method significantly reduces the time required for generating effective test cases,thereby enhancing the efficiency of industrial control protocol vulnerability discovery.

Research Progress of Backdoor Attacks in Deep Neural Networks

HUANG Shuxin, ZHANG Quanxin, WANG Yajie, ZHANG Yaoyuan, LI Yuanzhang

Computer Science. 2023, 50 (9): 52-61. doi:10.11896/jsjkx.230500235

Abstract

PDF(2662KB) ( 2072 )

References | Related Articles | Metrics

In recent years,deep neural networks(DNNs) have developed rapidly,and their applications involve many fields,including auto autonomous driving,natural language processing,facial recognition and so on,which have brought a lot of convenience to people's life.However,the growth of DNNs has brought some security concerns.In recent years,DNNs have been shown to be vulnerable to backdoor attacks,mainly due to their low transparency and poor interpretability,allowing attackers to to swoop in.In this paper,the potential security and privacy risks in neural network applications are revealed by reviewing the research work related to neural network backdoor attacks,and the importance of research in the field of backdoor is emphasized.This paper first briefly introduces the threat model of neural network backdoor,then the neural network backdoor attack is divided into two categories:the backdoor attack based on poisoning and the backdoor attack without poisoning,and the poisoning attack can be subdivided into multiple categories.It aggregates available resources about backdoor attack,and analyzes the development of backdoor on neural network and the future development trend of backdoor attack is prospected.

Privacy-enhanced Federated Learning Algorithm Against Inference Attack

ZHAO Yuhao, CHEN Siguang, SU Jian

Computer Science. 2023, 50 (9): 62-67. doi:10.11896/jsjkx.220700174

Abstract

PDF(2843KB) ( 1885 )

References | Related Articles | Metrics

In federated learning,each distributed client does not need to transmit local training data,the central server jointly trains the global model by gradient collection,it has good performance and privacy protection advantages.However,it has been demonstrated that gradient transmission may lead to the privacy leakage problem in federated learning.Aiming at the existing problems of current secure federated learning algorithms,such as poor model learning effect,high computational cost,and single attack defense,this paper proposes a privacy-enhanced federated learning algorithm against inference attack.First,an optimization problem of maximizing the distance between the training data obtained by inversion and the training data is formulated.The optimization problem is solved based on the quasi-Newton method to obtain new features with anti-inference attack ability.Second,the gradient reconstruction is achieved by using new features to generate gradients.The model parameters are updated based on the reconstructed gradients,which can improve the privacy protection capability of the model.Finally,simulation results show that the proposed algorithm can resist two types of inference attacks simultaneously,and it has significant advantages in protection effect and convergence speed compared with other secure schemes.

Study on Dual-security Knowledge Graph for Process Industrial Control

WANG Jing, ZHANG Miao, LIU Yang, LI Haoling, LI Haotian, WANG Bailing, WEI Yuliang

Computer Science. 2023, 50 (9): 68-74. doi:10.11896/jsjkx.230500233

Abstract

PDF(3170KB) ( 1989 )

References | Related Articles | Metrics

With the development of industrial control systems,security issues in these systems have become increasingly important.However,traditional industrial safety systems usually focus on either information security or production safety,thus failing to consider both issues at the same time.As structured representation of data,knowledge graph(KG) is capable of hosting domain-specific knowledge and modeling causal relationships among knowledge.However,most studies leverage KG to handle cybersecurity,while rarely pay attention to information security and production safety problems in industrial control systems.This paper proposes a set of construction methods for dual-security KG for process industrial control systems.Using the techniques of named entity recognition and relation extraction,it builds a large number of dual-security knowledge triples from a real-world production corpus.The built KG incorporates both features of chemical industry production process and potential network security flaws,providing comprehensive security guarantee for industrial control system.

Tiny Person Detection for Intelligent Video Surveillance

YANG Yi, SHEN Sheng, DOU Zhiyang, LI Yuan, HAN Zhenjun

Computer Science. 2023, 50 (9): 75-81. doi:10.11896/jsjkx.230400204

Abstract

PDF(1861KB) ( 2106 )

References | Related Articles | Metrics

Person detection has significant practical implications for social governance and urban security.Monitoring data is an important source of data security.Tiny object detection,which focuses on less than 20 pixels objects in large-scale images,is a challenging task.One of the main challenges is the scale mismatch between the dataset used for pre-training/co-training the detectors,such as COCO,and the dataset used for fine-tuning the detectors,such as TinyPerson,which negatively affects the performance of detectors on tiny object detection.To address this challenge,this paper proposes an optimization strategy called scale distribution searching(SDS) to match the scale of different datasets for tiny object detection,which also balance the information gain and loss.The Gauss model is used to model the scale distribution of targets in the dataset,and the optimal distribution parameters are found through iteration.The feature distribution and the performance of the detector is comparedto find the best scale distribution.Through the SDS strategy,mainstream object detection methods have achieved better performance on TinyPerson,demonstrating the effectiveness of the SDS strategy in improving pre-training/co-training efficiency.

Microservice Moving Target Defense Strategy Based on Adaptive Genetic Algorithm

LIU Xuanyu, ZHANG Shuai, HUO Shumin, SHANG Ke

Computer Science. 2023, 50 (9): 82-89. doi:10.11896/jsjkx.221000199

Abstract

PDF(2028KB) ( 1877 )

References | Related Articles | Metrics

Microservice architecture can effectively improve the agility of software due to its flexible,scalable and other characte-ristics,and has become the most mainstream method of application delivery in the cloud.However,the microservice splitting makes the attack surface of applications grow explosively,which brings great challenges to the design of mobile target defense strategy with the core of “strategic defense”.To solve this problem,a microservice moving target defense strategy based on adaptive genetic algorithm(AGA),namely dynamic rotation strategy(DRS),is proposed.Firstly,based on the characteristics of microservice,the attack path of attackers is analyzed.Then,a microservice attack graph model is proposed to formalize various attack scena-rios,and the security gains and return of defense(RoD) of moving target defense strategies are quantitatively analyzed.Finally,AGA is used to solve the optimal security configuration of mobile target defense,that is,the optimal dynamic rotation cycle of microservices.Experiments show that DRS is scalable,and the defense return rate of DRS increases by 17.25%,41.01% and 222.88% respectively compared with the unified configuration strategy,DSEOM and random configuration strategy.

Active Observation Schemes and Software Implementation Architecture of Autonomous Robot

XIAO Huaiyu, YANG Shuo, MAO Xinjun

Computer Science. 2023, 50 (9): 90-100. doi:10.11896/jsjkx.221200053

Abstract

PDF(4588KB) ( 3912 )

References | Related Articles | Metrics

Autonomous robots operate in the open environment,their perception of environmental information is limited,and it is difficult to obtain complete and timely information about the environment.In order to effectively complete tasks,autonomous robots need to actively observe the environment,that is,according to the requirements of tasks,to make decisions,schedule and execute observation behaviors spontaneously,and obtain task-related environmental information.The demand for active observation poses two challenges to the observation scheme of autonomous robots and the construction of software systems.On the one hand,in order to support the effective implementation of tasks,active observation schemes should be designed to ensure that autonomous robots can observe the required environmental information based on task requirements from the mechanism level.On the other hand,the active observation schemes make the function abstraction and data interaction of the software components such as observation and decision-making of autonomous robots more complicated,so it is necessary to design a software architecture sui-table for the implementation of the complex mechanism on the upper level.In order to deal with the above challenges,this paper defines the behaviors of autonomous robots as task behaviors and observation behaviors.Two kinds of active observation schemes are proposed to construct a collaborative mechanism between observation behaviors and task behaviors,aiming at the two typical scenes with limited environmental information in the open environment:one-sided observation and outdated observation scenes,and the decision and scheduling algorithms of observation behavior are designed based on these two active observation schemes.In addition,an autonomous robot software architecture based on the multi-agent system is designed to implement the proposed active observation schemes.Finally,in order to verify the effectiveness of the proposed active observation schemes,a typical task in the open environment:the book transfer task of the library service robot is selected to carry out experimental verification.In this task,the location information of the book is limited by the autonomous robot,which easily leads to the failure of the book transfer task.In this paper,the reactive observation and the accompanying observation schemes of the current mainstream in the field of autonomous robots are selected as the comparison method,and the effectiveness of the proposed method is verified by comparing the behavior execution process,motion trajectory and time consumption.

Study on REST API Test Case Generation Method Based on Dependency Model

LIU Yingying, YANG Qiuhui, YAO Bangguo, LIU Qiaoyun

Computer Science. 2023, 50 (9): 101-107. doi:10.11896/jsjkx.220800071

Abstract

PDF(2740KB) ( 3996 )

References | Related Articles | Metrics

The prevalence of dependencies in REST API makes it difficult to generate a reasonable sequence of API calls with input parameters.Most existing approaches only consider one of these dependencies and require cumbersome manual preliminaries,thus the generated test cases are still less effective.To address the above problem,a test case generation method based on depen-dency model is proposed.By parsing the OpenAPI documentation,this method extracts the inter-operation dependencies and inter-parameter dependencies,establishes two dependency models,generates test cases from the models,and determines test oracles.Experimental results show that the proposed method achieves 100% input metric coverage,and 100%,91.67%,and 83.33% coverage for status code category,status code,and response resource type,respectively,and can detect internal interface defects within a limited time.Compared with RESTler and RESTest,the proposed method improves the maximum 36% of output metric coverage,triggeres the most number of abnormal response status codes,and detects a maximum of 10% increase in the percentage of abnormal responses.The method provides a valuable reference for the test case generation problem of REST API.

Termination Analysis of Single Path Loop Programs Based on Iterative Trajectory Division

WANG Yao, LI Yi

Computer Science. 2023, 50 (9): 108-116. doi:10.11896/jsjkx.220700214

Abstract

PDF(1693KB) ( 3762 )

References | Related Articles | Metrics

The ranking function has been extensively studied as an important method of program termination analysis.In this paper,we focus on the termination of single-path loops.Firstly,the concept of two-way iterative loops is proposed,and the single-path loops are divided into bidirectional iterative loops and non-bidirectional iterative loops.Secondly,for the bidirectional iterative loop program,a division method and concept of trivial ranking function are proposed,and it is proved that if a bidirectional iterative loop exists a trivial rank function,it is terminated.As for the non-bidirectional iterative loop,we use incremental function as the division method,i.e.,the original program space is divided into smaller spaces by using incremental function,and the termination of the original program is proved by computing the ranking function on the smaller space.Finally,the problem of computing the trivial ranking function comes down to the SVM classification problem,and we verifies candidate ran-king functions using the tools Z3 or bottema.

Smart Contract Fuzzing Based on Deep Learning and Information Feedback

ZHAO Mingmin, YANG Qiuhui, HONG Mei, CAI Chuang

Computer Science. 2023, 50 (9): 117-122. doi:10.11896/jsjkx.220800104

Abstract

PDF(2388KB) ( 3825 )

References | Related Articles | Metrics

Vulnerabilities of smart contracts caused by insecure programming have been frequently discovered on the mainstream blockchain platform Ethereum.In order to improve the coverage of contracts by fuzzing and detect security vulnerabilities more comprehensively,this paper proposes a smart contract fuzzing.First,constructing Ethereum smart contract transaction sequence data set,then building smart contract generation model based on deep learning to generate initial seeds for fuzzing.Then,accor-ding to the information of coverage and branch distance,conduct information feedback-guided fuzzing on smart contracts,a speci-fic chromosome encoding method for test cases is proposed,and corresponding crossover operators and mutation operators are designed and implemented.The method can effectively cover the deep state of smart contracts and branch code guarded by strict conditions.Experiments on 500 smart contracts show that the code coverage rate of this method is 93.73%,and the vulnerability detection rate is 93.93%.Compare with the ILF,sFuzz,and Echidna methods,the code coverage rate of this method increases by 3.80%～25.49%,the vulnerability detection rate increases by 4.64%~24.02%.This method helps to improve the effectiveness of Ethereum smart contract security testing and is worthy of reference for the industry.

Contrastive Clustering with Consistent Structural Relations

XU Jie, WANG Lisong

Computer Science. 2023, 50 (9): 123-129. doi:10.11896/jsjkx.220700288

Abstract

PDF(2489KB) ( 643 )

References | Related Articles | Metrics

As a basic unsupervised learning task,clustering aims to divide unlabeled and mixed images into semantically similar classes.Some recent approaches focus on the ability of the model to discriminate between different semantic classes by introducing data augmentation,using contrastive learning methods to learn feature representations and cluster assignments,which may lead to situations that feature embeddings from samples with the same semantic class are separated.Aiming at the above problems,a comparative clustering method with consistent structural relations(CCR) is proposed,which performs comparative learning at the instance level and cluster level,and adds consistency constraints at the relationship level.So that the model can learn more information of ‘positive data pair' and reduce the impact of cluster embedding being separated.Experimental results show that CCR obtains better results than the unsupervised clustering methods in recent years on the image benchmark dataset.The average accuracy on the CIFAR-10 and STL-10 datasets improves by 1.7% compared to the best methods in the same experimental settings and improves by 1.9% on the CIFAR-100 dataset.

Graph Similarity Search with High Efficiency and Low Index

QIU Zhen, ZHENG Zhaohui

Computer Science. 2023, 50 (9): 130-138. doi:10.11896/jsjkx.220700105

Abstract

PDF(2297KB) ( 639 )

References | Related Articles | Metrics

Graph similarity search is to search the graph set that is similar to query graph under a measurement,which adopts the “filtering-verification” framework.Aiming at the problems of the existing methods,such as the untight lower bound and the large index space,an improved graph similarity search algorithm(Z-Index) based on query graph partition with multi-level filtering and low index space is proposed.Firstly,the pre-candidate set is obtained by global coarse-grained filtering.Secondly,a query graph partitioning algorithm based on extension probability is proposed,and a hierarchical filtering mechanism is adopted to further shrink the candidate set,so as to enhance the tightness of the lower bound.Finally,the sequence similarity difference is introduced to compute the sparsity of the data contribution.Then partition compression and difference compression algorithm are proposed to construct “zero” index structure,so as to reduce the index space.Experimental results show that Z-Index algorithm has a tighter lower bound,and the candidate set size of Z-Index can be reduced about 50%.Moreover,the algorithm execution time is greatly reduced,and it still shows great scalability in the case of tiny index space.

Human Mobility Pattern Prior Knowledge Based POI Recommendation

YI Qiuhua, GAO Haoran, CHEN Xinqi, KONG Xiangjie

Computer Science. 2023, 50 (9): 139-144. doi:10.11896/jsjkx.220900114

Abstract

PDF(2574KB) ( 603 )

References | Related Articles | Metrics

Point of interest(POI) recommendation is a fundamental task in location-based social networks,which provides users with personalized place recommendations.However,the current point of interest recommendation is mostly based on learning the user's check-in history at the point of interest in the social network and the user relationship network for recommendation,and the travel rules of urban crowds cannot be effectively used.To solve the above problem,firstly,a human mobility pattern extraction(HMPE) framework is proposed,which takes advantage of graph neural network to extract human mobility pattern.Then attention mechanism is introduced to capture the spatio-temporal information of urban traffic pattern.By establishing downstream tasks and designing upsampling modules,HMPE restores representation vectors to task objectives.An end-to-end framework is built to complete pre-training of human mobility pattern extraction module.Secondly,the human mobility tecommendation(HMRec)algorithm is proposed,which introduces the prior knowledge of crowd movement patterns,so that the recommendation results are more in line with human travel intentions in cities.Extensive experiments show that HMRec is superior to baseline mo-dels.Finally,the existing problems and future research directions of interest point recommendation are discussed.

Study on Supervised Learning Model for Optimal Histogram Solution

CHEN Yunliang, LIU Hao, ZHU Guishui, HUANG Xiaohui, CHEN Xiaodao, WANG Lizhe

Computer Science. 2023, 50 (9): 145-151. doi:10.11896/jsjkx.230300065

Abstract

PDF(1675KB) ( 640 )

References | Related Articles | Metrics

The dynamic programming binning algorithm is currently used to realize the optimal histogram.However,its time complexity is too high.A supervised learning model based on ProbSparse self-attention is proposed in this paper to learn the dynamic programming binning algorithm.It can be used as an alternative to the dynamic programming binning algorithm.The proposed model consists of three parts:1)mapping the numerical input sequence into the corresponding vector sequence through the embedding and position coding layer;2)capturing the dependence between input sequences through the ProbSparse self-attention;3)the dependency is mapped to the subscript information of the binning “bucket” boundary through the feedforward neural network.Experimental results on six data sets indicate that the proposed model based on ProbSparse self-attention outperforms the dynamic programming binning algorithm.The accuracy of the proposed method is greater than 83.47%.Meanwhile,its time cost in the prediction stage is no more than 1/3 of the compared method.

Rectifying Dual Bias for Recommendation

HUANG Lu, NI Lyu, JIN Cheqing

Computer Science. 2023, 50 (9): 152-159. doi:10.11896/jsjkx.220900035

Abstract

PDF(2466KB) ( 577 )

References | Related Articles | Metrics

In recent years,a large number of recommendation algorithms have emerged,most of which focus on how to construct a machine learning model to give a good fit to historical interaction data.However,historical interaction data always come from observations rather than experiments in recommendation.Various biases exist in observed data,where the popularity bias is a representative one.Most approaches to dealing with popularity bias use the strategy of removing the popularity bias.But it is actually difficult for these approaches to improve the recommendation accuracy due to bias amplification causedby recommendation algorithms.Thus,the strategy of leveraging the popularity bias bothin training and inferencestagesis more applicable.Combined with the causal graph,a double bias deconfounding and adjusting(DBDA) model is proposed to rectify bias from the perspectives of both user and item.In the training stage,the adverse effects of the popularity bias are removed,and in the inference stage,a more accurate prediction of user preferences is made with the aid of the trend of popularity.Experiments are conducted on three large-scale public datasets to verify that the proposed method produces 2.48%~19.70% higher diverse evaluation metrics than the state-of-art method.

Multi-task Graph-embedding Deep Prediction Model for Mobile App Rating Recommendation

LI Haiming, ZHU Zhiheng, LIU Lei, GUO Chenkai

Computer Science. 2023, 50 (9): 160-167. doi:10.11896/jsjkx.220700035

Abstract

PDF(2217KB) ( 634 )

References | Related Articles | Metrics

With the prevalence of smart terminal devices and mobile application(app for short),the requirements for application quality and user experience gradually increase.As an effective pre-assessment method,mobile app rating recommendation has gained increasing attention from app markets.The traditional app rating and recommendation works mainly focus on challenges such as data sparsity and model depth.Nevertheless,they fail to accurately capture the graph relationship within the apps and users.Furthermore,the multi-task characteristic of app recommendation is neglected.Aiming at these shortcomings,this paper proposes a graph embedding multi-task model AppGRec for mobile app rating and recommendation.AppGRec uses the embedding structure of inductive bipartite graph to mine the user interaction features.It uses the shared-bottom based model to capture the multi-task feature in app rating,while considering the effects of data sparsity and model depth.16 031 valid mobile apps and their feature data on Google Play are collected as dataset for method evaluation.Experimental results show that AppGRec achieves 10.4% and 10.9% improvement in terms of MAE and RMSE respectively comparing with the state-of-the-art models.In addition,this paper also makes quantitative analysis of the impact of hyperparameters and some core modules in AppGRec,and verifies the effectiveness from multiple perspectives.

Image Relighting Network Based on Context-gated Residuals and Multi-scale Attention

WANG Wei, DU Xiangcheng, JIN Cheng

Computer Science. 2023, 50 (9): 168-175. doi:10.11896/jsjkx.221000100

Abstract

PDF(4180KB) ( 610 )

References | Related Articles | Metrics

Image relighting is commonly used in image editing and data augmentation tasks.Existing image relighting methods suffer from estimating accurate shadows and obtaining consistent structures and clear texture when removing and rendering sha-dows in complex scenes.To address these issues,this paper proposes an image relighting network based on context-gated resi-duals and multiscale attention.Contextual gating residuals capture the long-range dependencies of pixels by aggregating local and global spatial context information,which maintains the consistency of shadow and lighting direction.Besides,gating mechanisms can effectively improve the network's ability to recover textures and structures.Multiscale attention increases the receptive field without losing resolution by iteratively extracting and aggregating features of different scales.It activates important features by concatenating channel attention and spatial attention,and suppresses the responses of irrelevant features.In this paper,lighting gradient loss is also proposed to obtain satisfactory visual images through efficiently learning the lighting gradients in all directions.Experimental results show that,compared with the current state-of-the-art methods,the proposed method improves PSNR and SSIM by 7.47% and 12.37%,respectively.

Infrared Ground Multi-object Tracking Method Based on Improved ByteTrack Algorithm

WANG Luo, LI Biao, FU Ruigang

Computer Science. 2023, 50 (9): 176-183. doi:10.11896/jsjkx.220900004

Abstract

PDF(3455KB) ( 754 )

References | Related Articles | Metrics

The research of infrared object intelligent detection and tracking technology is always a hot topic in the same field,especially in precision guidance,sea surface surveillance and sky warning.Aiming at the problems that tracking accuracy is reduced due to ground miscellaneous interference,multi-object block interference,platform shaking and other complex scenes,an infrared ground multi-object tracking method based on improved ByteTrack algorithm is proposed.First of all,a modified Kalman filter which could adaptively modulate the noise scale is introduced to alleviate the impact of low-quality detection on vanilla Kalman filter.Secondly,the introduction of enhanced correlation coefficient maximization is used to settle the inter-frame images to compensate the platform shaking impact.Then ByteTrack increases the motion model based on long short-term memory network to solve the prediction error caused by Kalman filter in the non-linear motion state.Finally,two lightweight offline algorithms of link model and Gaussian-smoothed interpolation are introduced to refine the tracking results.Experiment is performed on the infrared ground multi-object dataset and the results show that compared with Sort and DeepSort,the MOTA of the improved algorithm increases by 8.3% and 10.2%,IDF1 increases by 6.5% and 5.6%,respectively.The improved algorithm shows better effectiveness and will be used in the infrared object intelligent detection and tracking scenes.

Sign Language Animation Splicing Model Based on LpTransformer Network

HUANG Hanqiang, XING Yunbing, SHEN Jianfei, FAN Feiyi

Computer Science. 2023, 50 (9): 184-191. doi:10.11896/jsjkx.221100043

Abstract

PDF(3330KB) ( 620 )

References | Related Articles | Metrics

Sign language animation splicing is a hot topic.With the continuous development of machine learning technology,especially the gradual maturity of deep learning related technologies,the speed and quality of sign language animation splicing are constantly improving.When splicing sign language words into sentences,the corresponding animation also needs to be spliced.Traditional algorithms use distance loss to find the best splicing position when splicing animation,and use linear or spherical interpolation to generate transition frames.This splicing algorithm not only has obvious defects in efficiency and flexibility,but also gene-rates unnatural sign language animation.In order to solve the above problems,LpTransformer model is proposed to predict the splicing position and generate transition frames.Experiment results show that the prediction accuracy of LpTransformer's transition frames reaches 99%,which is superior to ConvS2S,LSTM and Transformer,and its splicing speed is five times faster than Transformer,so it can achieve real-time splicing.

Super Multi-class Deep Image Clustering Model Based on Contrastive Learning

HU Shen, QIAN Yuhua, WANG Jieting, LI Feijiang, LYU Wei

Computer Science. 2023, 50 (9): 192-201. doi:10.11896/jsjkx.220900133

Abstract

PDF(2493KB) ( 669 )

References | Related Articles | Metrics

Image clustering reduces the dimensionality of image data,extracts effective features through representation learning,and performs cluster analysis.When there are many categories of image data,the complexity of data distribution and the density of clusters seriously affect the practicability of existing methods.To this end,this paper proposes a super-multi-class deep image clustering model based on contrastive learning,which is mainly divided into three stages:firstly,improving the contrastive lear-ning method to train the feature model to make the cluster distribution uniform;secondly,based on the principle of semantic similarity,the perspective mines instance semantic nearest neighbor information;and finally,the instance and its nearest neighbors are used as self-supervised information to train a clustering model.According to the different types of experiments,ablation experiments and contrast experiments are designed in this paper.The ablation experiments prove that the proposed method could make the clusters evenly distributed in the mapping space and mine the semantic nearest neighbor information reliably.In the comparative experiments,it's compared with the advanced algorithms on 7 benchmark datasets.On the ImageNet-200 class dataset,it's accuracy is 10.6% higher than the advanced method.It's accuracy rate on the ImageNet-1000 class dataset is higher than that of the advanced algorithm,which improves by 9.2%.

Study on Building Extraction Algorithm of Remote Sensing Image Based on Multi-scale Feature Fusion

CHEN Guojun, YUE Xueyan, ZHU Yanning, FU Yunpeng

Computer Science. 2023, 50 (9): 202-209. doi:10.11896/jsjkx.220800086

Abstract

PDF(4236KB) ( 670 )

References | Related Articles | Metrics

Because of the various size of buildings and complicated background in high-resolution remote sensing images,there are some problems such as loss of details and blurring of edges when extracting buildings in remote sensing images,which affect the segmentation accuracy of the model.In order to solve these problems,this paper proposes a two-branch architecture network B2Net with spatial and semantic information branches.Firstly,the cross feature fusion module is provided in the semantic information branch to fully capture the context information to aggregate more multi-scale semantic features.Secondly,in the spatial branch,we combine the atrous convolution and depthwise separable convolution to extract the multi-scale spatial features of the image,and optimize the dilated rate to expand the receptive field.Finally,we use the content aware attention module to adaptively select the high-frequency and low-frequency content in the image to achieve the effect of refining the edges of building segmentation.We train and test the B2Net on two building datasets.On the WHU dataset,compared with the baseline model,the B2Net achieves the best result in precision,recall,F1 score and IoU,which is 98.60%,99.40%,99.30%,and 88.50%,respectively.On the Massachusetts building dataset,the four indicators are 0.9%,1.9%,1.7% and 2.2% higher than BiSeNet,respectively.Experiments show that B2Net can better capture spatial detail and high-level semantic information,improve the segmentation accuracy of buildings in complicated backgrounds,and meet the needs of rapid extraction of buildings.

Deep Artificial Correspondence Generation for 3D Point Cloud Registration

BAI Zhengyao, XU Zhu, ZHANG Yihan

Computer Science. 2023, 50 (9): 210-219. doi:10.11896/jsjkx.220700023

Abstract

PDF(2821KB) ( 583 )

References | Related Articles | Metrics

To address the challenging problems of point cloud registration in 3D reconstruction(e.g.,difficulty in finding corresponding points,etc.),this paper proposes a point cloud registration method based on cross-attention and artificial correspondence generation mechanism,Deep Artificial Correspondence Generation(DeepACG),by fully utilizing the geometric information of the source and target point clouds.Our method adopts a three-stage network model.The first stage is the deep feature encoding module,which exchanges and enhances the contextual and structural information between two unaligned point clouds using the cross-attention mechanism.The second stage is the artificial correspondence generation module,which synthesizes the artificial correspondences by weighting the soft mapping.The third one is the correspondence weighting and outlier filtering module,which assigns different weights to the correspondence pairs and rejects them with a small probability.Extensive experiments are conducted on both synthetic and real-world datasets.Our method achieves a registration recall of 92.61% on the real-world indoor dataset 3DMatch,and we execute unseen partial registration experiments on ModelNet40,reducing the root mean square error of the rotation matrix and translation vector to 0.016 and 0.000 09,respectively.Experimental results show that DeepACG has higher registration accuracy and robustness,and its alignment error is lower than that of the existing mainstream registration approaches.

Self-supervised Learning for 3D Real-scenes Question Answering

LI Xiang, FAN Zhiguang, LIN Nan, CAO Yangjie, LI Xuexiang

Computer Science. 2023, 50 (9): 220-226. doi:10.11896/jsjkx.220900256

Abstract

PDF(2208KB) ( 585 )

References | Related Articles | Metrics

Visual question answering(VQA)has gradually become one of the research hotspots in recent years.Most of the current question-answering research is 2D-image-based,often suffering from spatial ambiguity introduced by viewpoint changing,occlusion,and reprojection.In practice,human-computer interaction scenarios are often three-dimensional,yielding the demand for 3D-scene-based question answering.Existing 3D question answering algorithms have so far been able to perceive 3D objects and their spatial relationships,and can answer complex questions.However,point clouds represented by 3D scenes and the target questions belong to two different modalities,which are extremely difficult to align,leading to their unconspicuous related features are easy to be ignored.Aiming at this problem,this paper proposes a novel learning-based question answering method for realistic 3D scenes,called 3D self-supervised question answering(3DSSQA).Within 3DSSQA,a 3D cross-modal contrastive learning model(3DCMCL) is proposed to first align point-cloud data with question data globally for modality heterogeneity gap reduction,before mining related features between the two.In addition,a deep interactive attention(DIA) network is adapted to align 3D objects with keywords in a more fine-grained granularity,facilitating sufficient interactions between them.Extensive experiments on the ScanQA dataset demonstrate that 3DSSQA achieves an accuracy of 24.3% on the main EM@1 metric,notably surpassing state-of-the-art models.

Deep Learning Based Salient Object Detection in Infrared Video

ZHU Ye, HAO Yingguang, WANG Hongyu

Computer Science. 2023, 50 (9): 227-234. doi:10.11896/jsjkx.220700204

Abstract

PDF(2765KB) ( 647 )

References | Related Articles | Metrics

In the face of massive infrared video images with more and more complex background,the performance of the tradi-tional methods for salient object detection decreases significantly.In order to improve the performance of salient object detection in infrared images,this paper proposes a deep learning-based salient object detection model for infrared video,which mainly consists of a spatial feature extraction module,a temporal feature extraction module,a residual skip connection module and a pixel-wise classifier.First,the spatial feature extraction module is used to extract spatial saliency features from raw input video frames.Secondly,the temporal feature extraction module is used to obtain temporal saliency features and spatio-temporal coherence mo-deling.Finally,the spatial-temporal feature information and the spatial low-level feature information obtained by connecting the spatial module with the residual skip connection layer are sent into the pixel-wise classifier to generate the final salient object detection results.To improve the stability of the model,BCEloss and DICEloss are combined to train the network.The test is carried out on infrared video dataset OTCBVS and infrared video sequences with complex background.The proposed model can obtain accurate salient object detection results,and has robustness and good generalization ability.

Crowd Counting Based on Multi-scale Feature Aggregation in Dense Scenes

LIU Peigang, SUN Jie, YANG Chaozhi, LI Zongmin

Computer Science. 2023, 50 (9): 235-241. doi:10.11896/jsjkx.220800067

Abstract

PDF(3801KB) ( 649 )

References | Related Articles | Metrics

Individual scales vary greatly in dense scenes,and the varying scales of target individuals lead to poor crowd counting accuracy.To address this problem,the crowd counting method based on multi-scale feature fusion in dense scenes is proposed.The method investigates the ability of different feature layers to represent feature information for individuals at different scales,with adequate access to multi-scale features through layer connections.At the same time,a multi-scale feature aggregation module is proposed,which uses multiple columns of dilated convolution with different expansion rates,and automatically adjusts the perceptual field through a dynamic feature selection mechanism to effectively extract features of individuals at different scales.The method can further expand the field of perception while preserving the information of small-scale,and improving the detection capability of large-scale individuals,making it better adapted to the multi-scale changes of the population.Experimental results on the three public population counting datasets show that the proposed model has further improved the counting accuracy,with an MAE of 51.21 and an MSE of 83.70 on the ShanghaiTech Part A dataset.

Overview About Composite Semantic-based Event Graph Construction

ZHAI Lizhi, LI Ruixiang, YANG Jiabei, RAO Yuan, ZHANG Qitan, ZHOU Yun

Computer Science. 2023, 50 (9): 242-259. doi:10.11896/jsjkx.230400046

Abstract

PDF(4497KB) ( 1807 )

References | Related Articles | Metrics

The world is made up of countless interconnected events and the social activities of human beings are often driven by these various events.Research on the process of evolution and influence of events can not only helps us understand the evolution laws of human behaviors and social activities,but also provide a strategy for reasoning and thinking about artificial intelligence techniques,which has been paid a lot attention and becomes one of the new hottest research field.Unlike traditional knowledge graph,event graphs can abstract various events from the real world as nodes and recognize the logical relationships between events,such as state transforms or action sequences between different events,to form an innovation knowledge network with some composite semantic features.From the higher-level semantic viewpoints,the evolution of the complex events reflects the process of social activity with a certain of hidden logical relationships behind of them.In this paper,some critical challenges in the process of event graph construction have been analyzed,i.e.,how to extract the event in open domain,to establish a common event standards,to extract the relationship between events,to fusion and optimize the event graph,and to build a strategy for event graph representation learning.In addition,this paper also overviews and summarizes some core technologies,public evaluation data sets,related measure indicators,and then some research directions in future have been illustrated.

Design of Visual Context-driven Interactive Bot System

LIU Yubo, GUO Bin, MA Ke, QIU Chen, LIU Sicong

Computer Science. 2023, 50 (9): 260-268. doi:10.11896/jsjkx.230200167

Abstract

PDF(3000KB) ( 1709 )

References | Related Articles | Metrics

Bots are intelligent software that can interact with people,and usually have the characteristics of real-time and interactivity.This paper takes the bots driven by visual context awareness as the theme,and explores from four aspects:lightweight target detection model and compression,real-time key frame extraction,system optimization,and interaction strategy,and builds strong real-time on edge resource-constrained devices.A flexible,highly interactive and highly scalable bots system.Specifically,in terms of lightweight target detection models and compression,we first explore the performance and accuracy of different lightweight target detection models,and compress the SSD model based on the VGG16 network to find a suitable compression strategy.Compression on the latest SSD model can increase the frame rate by 187% compared with the original model,under the pre-mise that the accuracy loss does not exceed 0.1%.In terms of real-time key frame extraction,the input video stream is pre-screened to reduce system pressure,which is equivalent to reducing inference delay by 90%.In terms of system optimization,the use of microservices reduces the cold start delay by about 98%.In terms of interaction strategy,a state machine with timer is used to model the situation to achieve situation-driven,and the output of human-computer interaction is completed in the form of speech.

TAMP:A Hierarchical Multi-robot Task Assignment Method for Area Coverage

AN Haojia, SHI Dianxi, LI lin, SUN Yixuan, YANG Shaowu, CHEN Xucan

Computer Science. 2023, 50 (9): 269-277. doi:10.11896/jsjkx.220800094

Abstract

PDF(3135KB) ( 1672 )

References | Related Articles | Metrics

As the foundation of many mobile robot applications,complete coverage aims to plan a collision-free path for robot to visit all points in the target area quickly.Using multiple robots for cooperative coverage can significantly reduce coverage time and improve system robustness.However,it increases the algorithm's complexity and makes cooperative robot management more challenging.Therefore,the multi-robot coverage problem in a given environment is studied in this paper,which has been proven to be an NP problem.This work proposes a heuristic multi-robot task assignment based on multi-level graph partitioning(TAMP) method,which consists of a coarse task assignment algorithm and a fine task assignment algorithm.The coarse task assignment algorithm reduces the size of the graph by the strategies of multi-level coarsening and graph maximal matching and then obtains a roughly balanced task assignment by the graph growth strategy.The fine task assignment algorithm proposes a Lazy&Lock strategy to achieve task subdivision,which improves the solution accuracy.Simulations validate the performance of the TAMP approach under different scales of random graphs and real-world policing patrol scenarios.Compared to the conventional task assignment method,TAMP expands the maximum computational scale from thousands to millions.For small-scale graphs(within 3 000),TAMP accelerates computation time by 20 times and outperforms the conventional method in terms of the deviation from the optimal solution for different small-scale graphs.For large-scale graphs(3 000~1 million),TAMP can solve the task assignment problem in 60 s while keeping the deviation of the optimal solution within 0.3%.

Hierarchical Multi-label Text Classification Algorithm Based on Parallel Convolutional Network Information Fusion

YI Liu, GENG Xinyu, BAI Jing

Computer Science. 2023, 50 (9): 278-286. doi:10.11896/jsjkx.221200133

Abstract

PDF(3202KB) ( 1761 )

References | Related Articles | Metrics

Natural language processing(NLP) is an important research direction in the field of artificial intelligence and machine learning,which aims to use computer technology to analyze,understand,and process natural language.One of the main research areas in NLP is to obtain information from textual content and automatically classify and label textual content based on a certain labeling system or standard.Compared to single-label text classification,multi-label text classification has the characteristic that a data element belongs to multiple labels,which makes it more difficult to obtain multiple categories of data features from textual information.Hierarchical classification of multi-label texts isa special category,whichdivides the information contained in the text into different category labeling systems,and each category labeling system has an interdependent hierarchical relationship.Therefore,the use of the hierarchical relationship in the internal labeling system to more accurately classify the text into corresponding labels becomes the key to solving the problem.To this end,this paper proposes a hierarchical classification algorithm for multi-label texts based on the fusion of parallel convolutional network information.First,the algorithm uses the BERT model for word integration in textual information,then it enhances the semantic features of textual information using a self-attention mechanism and extracts the features of textual data using different convolutional kernels.The multi-faceted semantic information of the text is more effectively used for the task of a hierarchical classification of multi-label texts by using a threshold-controlled tree structure to establish inter-node relationships between higher and lower bits.The results obtained on the Kanshan-Cup public dataset and the CI enterprise information dataset show that the algorithm outperforms TextCNN,TextRNN,FastTex and other comparative models in three evaluation measures,namely macro-precision,macro-recall,and micro F1 value,and has a better cascade multi-label text classification effect.

Chinese Medical Named Entity Recognition Method Incorporating Machine ReadingComprehension

LUO Yuanyuan, YANG Chunming, LI Bo, ZHANG Hui, ZHAO Xujian

Computer Science. 2023, 50 (9): 287-294. doi:10.11896/jsjkx.220900226

Abstract

PDF(2783KB) ( 1730 )

References | Related Articles | Metrics

Medical named entity recognition is the key to automatically build a large-scale medical knowledge base.However,medical entities are often nested,and it can not be recognized by the sequence labeling method.This paper proposes a Chinese medical named entity recognition method based on reading comprehension framework.It models the nested named entity recognition problem as a machine reading problem,uses BERT to establish the connection between the reading comprehension problem and medical text,and introduces a multi-head attention mechanism to strengthen the semantic connection between the problem and nested named entity,and finally uses two classifiers to predict the beginning and end positions of entities.This method achieves the best results with an F1-score of 67.65% when compared with the current five mainstream methods.Compared with the most classical BiLSTM-CRF,the F1-score improves by 7.17%,and the nested “symptom” entities increase by 16.81%.

Fusion of Semantic and Syntactic Graph Convolutional Networks for Joint Entity and Relation Extraction

HENG Hongjun, MIAO Jing

Computer Science. 2023, 50 (9): 295-302. doi:10.11896/jsjkx.220700041

Abstract

PDF(2926KB) ( 261 )

References | Related Articles | Metrics

Entity and relation extraction task is the core task of information extraction.It plays an irreplaceable role in effectively extracting key information from explosive growth data,and is also the basic task of building a large-scale knowledge graph.Therefore,the research on entity relationship extraction task is of great significance for various natural language processing(NLP) tasks.Although the existing entity and relation extraction based on deep learning method has a very mature theory and good performance,there are still some problems,such as error accumulation,entity redundancy,lack of interaction,entity and relation overlap.Semantic information and syntactic information play an important role in NLP tasks.In order to make full use of them to solve the above problems,a fusion of semantic and syntactic graph convolutional networks binary tagging framework for relation triple extraction(FSSRel) is proposed.The model is divided into three stages.In the first stage,the start and end positions of the triple body are predicted.In the second stage,semantic features and syntactic features are extracted by semantic graph neural network and syntactic graph neural network respectively,and fused into the coding vector.In the third stage,the object position of each relation of the statement is predicted and marked to complete the extraction of the final triple.Experimental results show that the F1 value of the model increases by 2.5% and 1.6% respectively compared with the baseline model on the NYT dataset and the WebNLG dataset,and it also performs well on complex data with overlapping triples and multiple triples.

Improved K2 Algorithm Based on Two-step Search Strategy

XU Miao, WANG Huiling, LIANG Yi, QI Xiaolong, GAO Yang

Computer Science. 2023, 50 (9): 303-310. doi:10.11896/jsjkx.220700253

Abstract

PDF(2119KB) ( 225 )

References | Related Articles | Metrics

Bayesian network receiving increasing attention from researchers because of its strong uncertainty reasoning ability and causal representability.Learning a Bayesian network structure from data is an NP-hard problem.Among them,for the problem that the K2 algorithm strongly depends on the topological order of variables,an improved K2 learning method TSK2 is proposed,which combines variable neighbor sets and v-structure information.The proposed method effectively reduces the search scale in the order space and avoids prematurely falling into local optimum.Specifically,inspired by the orientation rules of the constraint algorithm,the method reliably adjusts the in-order position of the neighbors of the sink with the help of the identified v-structure and neighbor set information.Secondly,inspired by the basic structure of the shell net,with the help of the variable neighbor set information,the optimal sequence is obtained by performing the search of the three basic structures of shun-connection,sub-connection,and confluence-connection to accurately correct the order positions of parent nodes and child nodes.Experimental results show that the accuracy of the proposed algorithm is significantly improved compared with the comparison methods on the Asia and Alarm network datasets.A more accurate network structure can be learned.

Edge Intelligent Sensing Based UAV Space Trajectory Planning Method

LIU Xingguang, ZHOU Li, ZHANG Xiaoying, CHEN Haitao, ZHAO Haitao, WEI Jibo

Computer Science. 2023, 50 (9): 311-317. doi:10.11896/jsjkx.220800032

Abstract

PDF(3464KB) ( 1799 )

References | Related Articles | Metrics

With the emergence of a large number of frequency-using equipment,the radio environment for UAVs to perform tasks has become more and more complex,which puts forward higher requirements for UAVs to recognize the situation and autonomous obstacle avoidance.In view of this,this paper proposes a 3D trajectory planning method for UAVs based on side-end colla-boration.First,a UAV trajectory planning framework with side-end collaboration is proposed,which can synergistically improve the environment perception and autonomous obstacle avoidance capabilities of UAVs under communication connectivity constraints.Second,it proposes an artificial potential field method based on the deep deterministic policy gradient(DDPG) algorithm to avoid UAVs from falling into local minimum points and optimize UAV flight energy consumption.Finally,by performing simulation experiments in static and dynamic interference environments,compared with other trajectory planning methods,the proposed method can optimize the UAV flight trajectory and transmission data rate,which reduces the flight energy consumption of UAVs 5.59% and 11.99% respectively,and improve the transmission data rate 7.64% and 16.52% in static and dynamic interference environments.The proposed method also significantly improves the communication stability and the adaptability of UAVs to complex electromagnetic environments.

EGCN-CeDML:A Distributed Machine Learning Framework for Vehicle Driving Behavior Prediction

LI Ke, YANG Ling, ZHAO Yanbo, CHEN Yonglong, LUO Shouxi

Computer Science. 2023, 50 (9): 318-330. doi:10.11896/jsjkx.221000064

Abstract

PDF(3439KB) ( 1744 )

References | Related Articles | Metrics

In large-scale dynamic traffic scenarios,predicting vehicle driving behavior quickly and accurately is one of the most challenging issues in the field of intelligent traffic driving.The prediction of vehicle driving behavior should consider not only the efficiency of communication,but also the historical vehicle trajectory and the interaction between vehicles.Considering the above factors,this paper proposes a communication-efficient distributed machine learning framework based on edge-enhanced graph convolutional neural networks(EGCN-CeDML).Compared with the centralized prediction framework on a single device,EGCN-CeDML is a communication-efficient distributed machine learning framework,which does not need to transmit all the raw data to the cloud server,and directly stores,processes,and computes user data locally.This way of training neural networks on multiple edge devices relieves the pressure of centralized training neural networks,reduces the amount of transmitted data and communication latency,improves data processing efficiency,and preserves user privacy to a certain extent.EGCN-LSTM deployed on each edge device utilizes the edge-enhanced attention mechanism and the feature transfer mechanism of the graph convolutional neural network to promptly extract and transfer the interaction information between vehicles when the number of surrounding vehicles increases to more than a dozen,ensuring more accurate prediction performance and lower time complexity.In addition to vehicle driving behavior prediction,each edge device can flexibly control the type and scale of the neural network according to its own computing and storage capabilities,under the premise of ensuring the performance of the neural network,which is suitable for different application scenarios.The experimental results of EGCN-CeDML on public dataset NGSIMshow that the amount of data to be transmitted by only accounts for 21.56% of the centralized training.And the calculation time and prediction performance of EGCN-CeDML are better than those of previous models regardless of traffic complexity,with an accuracy rate of 0.939 1,a recall rate of 0.955 7,and an F1 score of 0.947 3.When the prediction time is one second,the prediction accuracy reaches 91.21%.Even if the number of vehicles increases,the algorithm maintains a low time complexity and is stable within 0.1 seconds.

Feature Weight Perception-based Prediction of Virtual Network Function Resource Demands

WANG Huaiqin, LUO Jian, WANG Haiyan

Computer Science. 2023, 50 (9): 331-336. doi:10.11896/jsjkx.221000012

Abstract

PDF(2441KB) ( 1697 )

References | Related Articles | Metrics

Virtual network function(VNF) provides services in the form of service function chain(SFC) to meet the performance requirements of different services.Due to the dynamic nature of the network,allocating fixed resources to VNF instances will lead to excessive or insufficient resources for VNF instances.Previous studies have not distinguished the importance of network load characteristics related to VNF profiles.Therefore,a dynamic VNF resource demand prediction method based on feature weight perception is proposed.Firstly,ECANet is used to learn the weight values of VNF features,to reduce the negative impact of useless features on the model prediction results.Secondly,because the VNF profile data set has structural characteristics,when building the VNF resource prediction model,it is necessary to consider mining the deep interrelationship between features by strengthening feature interaction.It is proposed to use the deep feature interactive network(DIN) to enhance the interaction between network load features and VNF performance features,so as to improve the prediction accuracy of the model.Finally,compared with similar methods on the benchmark dataset,it is found that the proposed method has more advantages in the effectiveness and accuracy of prediction.

Routing Protection Scheme with High Failure Protection Ratio Based on Software-defined Network

GENG Haijun, WANG Wei, ZHANG Han, WANG Ling

Computer Science. 2023, 50 (9): 337-346. doi:10.11896/jsjkx.220900220

Abstract

PDF(2195KB) ( 1706 )

References | Related Articles | Metrics

SDN has attracted extensive attention from academia for its advantages of strong programmability and centralized control.Existing SDN devices still use the shortest path protocol when performing packet forwarding.When a node in the shortest path fails,the network re-convergence is still required.During this period,packets may be discarded and thus cannot be delivered to the destination node,which has an impact on the flow of real-time applications and affects the user experience.The academia generally adopts the routing protection schemes to deal with network failures.The existing routing protection schemes have the following two problems:(1)the failure protection ratio is low;(2)when the network fails,the backup path may have routing loops.In order to solve the above two problems,a backup next hop calculation rule is proposed.Then,based on this rule,a routing protection algorithm with high hailure protection ratio(RPAHFPR) is designed,which combines the path generation algorithm(PGA),side branch first algorithm(SBF) and loop avoidance algorithm(LAA).It can simultaneously solve the low failure protection rate and routing loop problems faced by existing routing protection methods.Finally,the performance of RPAHFPR scheme is verified in a large number of real network topologies and simulated network topologies.Compared with the classic NPC and U-TURN,the failure protection rate of RPAHFPR is increased by 20.85% and 11.88% respectively,and it can achieve 100% fai-lure protection rate in 86.3% topology,and more than 99% failure protection rate in all topology.The path stretching degree of RPAHFPR is basically close to 1,without introducing too much time delay.

Task Offloading Algorithm Based on Federated Deep Reinforcement Learning for Internet of Vehicles

LIN Xinyu, YAO Zewei, HU Shengxi, CHEN Zheyi, CHEN Xing

Computer Science. 2023, 50 (9): 347-356. doi:10.11896/jsjkx.220800243

Abstract

PDF(2995KB) ( 1896 )

References | Related Articles | Metrics

With the rapid development of the service system of Internet of Vehicles applications,vehicles with limited computational resources have difficulty in handling these computation-intensive and latency-sensitive applications.As a key technique in mobile edge computing,task offloading can address the challenge.Specially,a task offloading algorithm based on federated deep reinforcement learning(TOFDRL) is proposed for dynamic multi-vehicle multi-road-side-unit(multi-RSU) task offloading environment in Internet of Vehicles.Each vehicle is considered as an agent,and a federated learning framework is used to train each agent.Each agent makes distributed decisions,aiming to minimize the average system response time.Evaluation experiments are set up to compare and analyze the performance of the proposed algorithm under a variety of dynamically changing scenarios.Si-mulation results show that the average response time of system solved by the proposed algorithm is shorter than that of the rule-based algorithm and the multi-agent deep reinforcement learning algorithm,close to the ideal scheme,and its solution time is much shorter than the ideal solution.Experimental results demonstrate that the proposed algorithm is able to solve an average system response time which is close to the ideal solution within an acceptable execution time.

Solution to Cross-domain VPN Based on Virtualization

TAO Zhiyong, ZHANG Jin, YANG Wangdong

Computer Science. 2023, 50 (9): 357-362. doi:10.11896/jsjkx.220800252

Abstract

PDF(3230KB) ( 1684 )

References | Related Articles | Metrics

To address the problems of complex implementation of cross-domain virtual private networks built in current carrier networks,excessive load on devices at the border of autonomous systems,and the existence of single points of failure,this paper proposes a solution for building cross-domain virtual private networks by virtualization.The scheme consists of four fundamental steps:the establishment of public network tunnels,the establishment of local VPN instances,the virtualization of autonomous system border devices,and the interaction of private network routes of border devices.To evaluate the feasibility and superiority of the scheme,comparative experiments are conducted with the cross-domain virtual private network constructed by the tradi-tional multi-hop EBGP approach in the dimensions of switching capacity,route entries,and label entries.Experimental results show that the cross-domain virtual private network constructed by this scheme enhances the data processing capability of the autonomous system boundary devices and reduces the amount of data to be processed by the autonomous system boundary devices.In general,this improved scheme is advanced and effective for building cross-domain virtual private networks.