Author Login Editor-in-Chief Peer Review Editor Work Office Work

15 November 2020, Volume 46 Issue 11
    

  • Select all
    |
    Hot Topics and Reviews
  • ZHANG Xiaorui, CHEN Xuan, SUN Wei, GE Kai
    Computer Engineering. 2020, 46(11): 1-11. https://doi.org/10.19678/j.issn.1000-3428.0058107
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    Vehicle re-recognition is a frontier and challenging subject in the field of computer vision,which aims at vehicle matching in non-overlapping field of view and multi-camera network.In recent years,deep learning technology has been successfully applied in vehicle re-identification tasks and seen as a research hotspot by virtue of its superior performance.To this,this paper expounds the research status of vehicle re-recognition based on deep learning,gives the definition of the vehicle re-recognition problem,and points out the limitations of the traditional vehicle re-identification methods and number-plate-based re-identification methods.Then this paper classifies and summarizes existing methods from different perspectives.By listing four commonly used vehicle re-identification data sets and comparing the performance of the classical methods on them,this paper provides a reference for the rational selection of vehicle re-identification methods in practical applications.On this basis,the challenges to the vehicle re-identification research are analyzed,and the development trend is prospected.
  • WANG Yihao, DING Hongwei, LI Bo, YANG Zhijun, YANG Jundong
    Computer Engineering. 2020, 46(11): 12-22. https://doi.org/10.19678/j.issn.1000-3428.0058802
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    The transmission of COVID-19 virus through respiratory droplets can be effectively prevented by correct mask wearing.However,complex factors in natural scenes including occlusion,crowds,and small-scale targets frequently affect the detection of mask wearing.To solve the problem,this paper proposes a YOLOv3-based mask wearing detection algorithm for complex scenes.The DarkNet53 backbone network is improved based on the cross-stage partial network,which reduces the calculation consumption and increases the training speed.Then an improved spatial pyramid pooling structure is introduced into YOLOv3,and the top-down and bottom-up feature fusion strategies are used to optimize the multi-scale prediction network,so as to realize feature enhancement.In addition,CIoU is selected as the loss function.The distance between the centers of the target and the detection frame,their overlap ratio,and aspect ratio are considered.The experimental results show that compared with the YOLOv3 algorithm,the proposed algorithm improves the detection accuracy of human faces by 7.3% and that of mask wearing by 14.9%,and the detection speed is improved by 6FPS on average.
  • ZHOU Jian, QU Ran
    Computer Engineering. 2020, 46(11): 23-28. https://doi.org/10.19678/j.issn.1000-3428.0057591
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    To address the security of the recovery of lost private keys on the blockchain,this paper proposes a distributed management scheme for private keys based on threshold secret sharing.The scheme combines the user’s private key with the secret password as a secret and uses the threshold key mechanism to divide the secret into several secret fragments,which are assigned to honest nodes for custody in the network by using the practical Byzantine fault tolerant algorithm.When the user’s private key is lost,the private key can be recovered by collecting secret fragments more than the threshold and combining them with the secret password.Analysis results show that this scheme is anonymous and resistant to collusion attacks and single node failure.It can implement recovery of the user’s lost private key while the dynamic management and secure storage of the private key is ensured.
  • YU Xiang, LIU Yixun, SHI Xueqin, WANG Zheng
    Computer Engineering. 2020, 46(11): 29-34,41. https://doi.org/10.19678/j.issn.1000-3428.0056850
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    In computing offloading systems of Mobile Edge Computing(MEC) in Internet of Vehicles(IoV),the loads of concurrent multi-priority computing tasks and MEC server resources are not balanced.To address the problem,this paper proposes a Genetic Algorithm-based Offloading Strategy(GAOS).The strategy sets weights of computing tasks with different priorities based on vehicle speed,MEC coverage and computing task features.On this basis,the computing tasks are coded.The problem of energy consumption minimization of computing offloading is transformed into a knapsack problem,and the optimal offloading strategy is obtained by using the genetic algorithm.Simulation results show that compared with the Random and all-MEC strategies,GAOS is least affected by the unbalanced loads of MEC servers,and increases the number of successfully processed on-board secure computing tasks by about 30% and 50% respectively.
  • SHU Fei, CHEN Tao, WANG Bin, YANG Huiting, LI Mingxuan
    Computer Engineering. 2020, 46(11): 35-41. https://doi.org/10.19678/j.issn.1000-3428.0057541
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    It is crucial to provide network security protection for power grid,one kind of national key infrastructure.Network security can be maintained by providing traffic warnings for the power grid Industrial Control System(ICS).To this end,this paper proposes an abnormal identification method that combines Deep Belief Network(DBN) with the Random Forest(RF) algorithm for the power grid ICS.The method constructs a DBN model to implement the in-depth mining of the correlation characteristics between multiple traffic characteristics and learn the feature extraction modes applicable to the traffic of power grid ICS.On this basis,the traffic whose characteristics are learnt and the malicious attack traffic are input into the RF detection model,and the model parameters are gradually adjusted for learning to obtain the optimal detection model.Tests are carried out on the data sets that are selected based on the features of power grid traffic from the classic intrusion detection data set,KDD99.Experimental results show that the accuracy rate of this method reaches 96.16% while the false alarm rate is only 3.49%.Compared with logistic regression model,multi-classification support vector machine model,DBN model and K-means algorithm,the proposed method can more accurately identify abnormal traffic in power grid ICS.
  • Artificial Intelligence and Pattern Recognition
  • HUANG Sheng, ZHANG Qianyun, LI Mengfang, ZHENG Xiufeng
    Computer Engineering. 2020, 46(11): 42-47. https://doi.org/10.19678/j.issn.1000-3428.0056282
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    In order to reduce the computational complexity of Screen Content Coding(SCC),this paper proposes a fast algorithm for SCC intra-frame CTU depth range prediction based on deep learning.A sufficient number of screen content video frame sequences are encoded as training data,and the distribution of CTU depth range is counted through a large amount of training data.The CTU category label is set according to the distribution ratio.Convolutional Neural Networks(CNN) architecture is designed and trained to predict the CTU depth range.Considering the CTU segmentation characteristics,the designed CNN architecture uses three different layers of convolution kernels to extract CTU depth-related features and provide training parameters for the CNN model.The trained CNN model is called at the time of encoding to predict the CTU depth range and reduce unnecessary depth traversal.Experimental results show that compared with SCM-8.0,the proposed algorithm saves an average of 48.34% coding time and increases the code rate by 2.59%,which effectively reduces the computational complexity of coding.
  • LUO Tiantian, ZHAO Lifeng
    Computer Engineering. 2020, 46(11): 48-52. https://doi.org/10.19678/j.issn.1000-3428.0056170
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    A hierarchical residual network constructed by the shortest augmented chain algorithm may lead to the loss of flow values when it faces multiple augmented chains with the same number of directed graphs and overlapped vertices,for it does not consider the augmented sequence when looking for augmented chains.In order to solve the problem,this paper proposes an improved network maximum flow algorithm with intersecting vertices in the network graph.It retains the hierarchical concept of the shortest augmented chain algorithm,and looks for augmented chains in the hierarchical residual network.On this basis,a rule is added to prioritize the search of the vertices that are related to the source vertex and with the minimum tolerance,and the vertices are taken as the new source vertices for further search.After an augmented chain is determined,the augmented chains with overlapped vertices are considered for further augmentation with the previous augmented chain.Results of example analysis and BA scale-free network modeling simulation show that,compared with the shortest augmented chain algorithm,the proposed algorithm can obtain more accurate maximum flow values while keeping same efficiency.
  • GAO Weijun, YANG Jie, ZHANG Chunxia, SHI Yang
    Computer Engineering. 2020, 46(11): 53-60. https://doi.org/10.19678/j.issn.1000-3428.0055950
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    Sentiment analysis is a fundamental field of Natural Language Processing(NLP).Convolutional Neural Network(CNN) performs well when applied to this field,but fails to fully extract the key sentiment information in the text information.To address the problem,this paper proposes a deep learning model,AT-DPCNN,which is based on attention mechanism.The model uses the attention matrix to highly focus on the part of the text sequence that has significant influence on emotion tendency,and operates on the extracted attention feature matrix and the word vector of the original text to get the attention input matrix.Then the CNN is used to re-extract text features.Also,in order to better extract the features of complex sentence patterns such as transition,the divide pooling is performed at the pooling layer.The proposed model is tested on different types of datasets,and the experimental results show that the model has good generalization performance,and significantly improves the classification accuracy and F1 score compared with WACNN,HAN and other models when processing complex sentence patterns such as transition.
  • TONG Manqi, HUANG Jiangsheng, GUO Kun
    Computer Engineering. 2020, 46(11): 61-69. https://doi.org/10.19678/j.issn.1000-3428.0056187
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    The speed of traditional user influence measurement algorithms is reduced when dealing with massive data.To address the problem,this paper proposes a comprehensive user influence measurement algorithm based on recessive interest.The Latent Dirichlet Allocation(LDA) model is used to obtain the recessive interests of the user,and the number of the optimal interest topics is determined based on the perplexity and the average topic similarity.Then,the transmission rate of user interests in the PageRank algorithm is improved to obtain the User Interest Factor(UIF).Finally,based on the Spark computing framework,the Analytic Hierarchy Process(AHP) is used to calculate the ultimate user influence by combining the influence of the user and UIF.Experimental results show that the proposed algorithm has a holistic consideration on user interests and the influence factors of the user,which enables it to provide more efficient and reasonable evaluation of the real influence of the user.
  • MA Zhekang, Diliyaer Paerhati, Zaokere Kadeer, Tuergen Yibulayin, Xerali Setti, Aishan Wumaier
    Computer Engineering. 2020, 46(11): 70-76. https://doi.org/10.19678/j.issn.1000-3428.0055990
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    To improve the utilization of key features of tourist question texts, this paper proposes a classification algorithm for tourist question texts integrated with the Word Level Convolutional Neural Network(WL-CNN) and the Sentence Level Bi-directional Long Short-Term Memory(SL-Bi-LSTM) network.The algorithm uses WL-CNN and SL-Bi-LSTM to learn the subspace vector of the word sequence and the deep semantic information of the sentence sequence.Then the two deep learning models are integrated by using the Multi-Head Attention Mechanism(MH-AM) to realize the syntactic and semantic information complementary of tourist question texts.Finally,the SoftMax classifier is used to obtain the classification results of tourist question texts.Experimental results show that the proposed algorithm has better classification performance than the tourist question text classification algorithms based on traditional deep learning models,increasing the accuracy to 0.986 6 and loss rate to 0.127 7.
  • WANG Yufeng, FENG Xinxi
    Computer Engineering. 2020, 46(11): 77-83. https://doi.org/10.19678/j.issn.1000-3428.0055586
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    To address the target tracking problem for multiple mobile sensor networks in the case of range-only measurement,this paper proposes a target tracking algorithm based on non-linear filtering and Multi-Dimensional Scaling(MDS) method.According to the relative motion between sensors and the target,a dynamic distance model with constraints is established.Then the Unscented Kalman Filtering(UKF) algorithm is used to improve the estimation precision of the distance as well as the rate of distance change in the model.On this basis,the position,the velocity and other state information of the sensors are used for calculation by the weighted MDS method to estimate the position and velocity of the target.Simulation results show that,when only the information of distance is accessible,the proposed algorithm can provide highly precise positioning for the target,as well as velocity estimation that can accurately reflect the real motion state of the target.Generally,the proposed algorithm performs better than ML-KF algorithm in target tracking.
  • CHENG Tao, CHEN Heng, LI Guanyu
    Computer Engineering. 2020, 46(11): 84-89. https://doi.org/10.19678/j.issn.1000-3428.0055720
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    Existing knowledge graph completion algorithms are time-consuming and inaccurate.To address these problems,this paper proposes a multi-layer convolution model based on half-edge.The model introduces the half-edge principle,and uses the descriptive information of the entity and the characteristics of the relation itself to constrain the head and tail entities connected by the relation based on their semantic similarity,so as to form the half-edge.On this basis,the Convolutional Neural Network(CNN) is used to complete the knowledge graph.In this model,the incomplete RDF triples containing only one entity and relationship are saved in the form of half-edge,which facilitates the completion of extended knowledge graphs and provides a foundation for the dynamic completion of knowledge graphs.Experimental results show that the proposed model has better performance in entity prediction and relationship prediction than TransE,DKRL and other models,and can effectively reduce the running time.
  • WANG Yiran, JING Xiaochuan, JIA Fukai, SUN Yujian, TONG Yi
    Computer Engineering. 2020, 46(11): 90-96. https://doi.org/10.19678/j.issn.1000-3428.0055904
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    There are multiple problems with existing multi-target tracking methods,including low learning speed,inefficient tracking and high difficulty in collaborative tracking strategy design.To this end,this paper proposes an improved multi-target tracking method.The method builds a task assignment model based on the number of target agents and tracking agents and their environmental information.Then the model is solved by using Hungary algorithm according to the distance benefit matrix to acquire the task assignment information of multiple tracking agents,which is optimized to shorten the tracking paths of target agents.In addition,the multi-agent collaborative reinforcement learning algorithm is used to enable multiple agents to repeat the process of exploration-accumulation-learning-decision in the same environment and update the strategy based on empirical data to finally complete the multi-target tracking task.Simulation results show that compared with DDPG and MADDPG methods,the proposed method enables multiple agents to collaboratively form the shortest path for tracking multiple moving targets with collisions and obstacles avoided.
  • YANG Shanshan, JIANG Lifen, SUN Huazhi, MA Chunmei
    Computer Engineering. 2020, 46(11): 97-103. https://doi.org/10.19678/j.issn.1000-3428.0055628
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    As a challenging task in the field of natural language processing,machine reading comprehension aims to answer questions related to articles and requires complex semantic reasoning.To solve the problem of information loss and inability to capture the global semantic relationship in feature extraction of existing machine reading comprehension methods,this paper constructs a multiple choice machine reading comprehension M-TCN model based on Temporal Convolutional Network(TCN).Attention mechanism is used to match articles,questions and candidate answers,and the internal relationship among them is established.At the same time,in order to extract high-level features to reduce information loss,TCN is used to aggregate the matching representation.The performance of the model is verified on the RCAE dataset of public reading comprehension.The experimental results show that compared with the existing machine reading comprehension models including ElimiNet,MRU and HCM,the proposed model increases the prediction accuracy to 52.5%,and its comprehensive performance is better.
  • QIN Tingting, LIU Zheng, CHEN Kejia
    Computer Engineering. 2020, 46(11): 104-108. https://doi.org/10.19678/j.issn.1000-3428.0055952
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    With the popularity of social software,mining effective information from massive digital documents has been a hotspot.The classic topic models including LDA and LSA capture topic information based on word co-occurrence and ignore the context information of words.To address the problem,this paper designs an attention mechanism between words and topics,integrates the topic information and word information into the LDA framework,and on this basis constructs a JEA-LDA topic model.The model uses the attention mechanism between words and topics to merge the word information and topic information into feature representation for topic extraction of the LDA model.The experimental results show that compared with LDA,DMM and other models,the proposed model has better performance in topic coherence and classification tasks,and improves the topic extraction results.
  • Advanced Computing and Data Processing
  • LI Jie, ZHU Hongliang, CHEN Yuling, XIN Yang
    Computer Engineering. 2020, 46(11): 109-116. https://doi.org/10.19678/j.issn.1000-3428.0056714
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    The Apriori algorithm can mine the association relationships between things,but the traditional Apriori algorithm needs to traverse the original transaction database every time the support of the candidate set is calculated,which reduces the efficiency of the algorithm.To address the problem,this paper proposes an improved algorithm based on hash storage and transaction weighting.The algorithm uses the deduplication feature of hash storage to deduplicate the transactions to reduce redundant calculations.At the same time,the mapping between the transaction set and the itemset is stored in the hash structure to avoid scanning the transaction database for multiple times during the calculation of the support of the candidate set.In addition,the support of the candidate set is calculated in parallel using multiple threads to improve the efficiency of the Apriori algorithm.Experimental results on open-source datasets show that the performance improvement of the proposed algorithm over the traditional Apriori algorithm increases with the number of transactions and repeated transactions in the dataset.Its running time is similar to that of the FP-Growth algorithm while the excessive memory consumption is avoided.
  • WU Weixin, HAN Jingyu, ZHU Man
    Computer Engineering. 2020, 46(11): 117-123. https://doi.org/10.19678/j.issn.1000-3428.0056311
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    With the development of semantic Web technology,the volume of Resource Description Framework(RDF) data is increasing rapidly along with its demand for storage space and transmission bandwidth.Existing general compression methods and RDF-specific compression methods can solve this problem,but still suffer from a lack of data redundancy.To this end,this paper proposes an RDF grouping compression algorithm based on delta encoding.The algorithm groups RDF data according to the combination of predicates connected to the object,so as to further reduce predicate redundancy while eliminating object redundancy.On this basis,it further optimizes the storage space of the grouped subject sequence data by introducing delta coding technology.Experimental results show that,compared with the Plain,HDT and HDT++ algorithm,this algorithm improves the performance by 17% on average in less structured datasets including Archives Hub,Linkedmdb,rdfabout and DBpedia,meanwhile improves performance by 23% on average in highly structured dataset dbtune,which demonstrates that the proposed algorithm has better RDF compression performance in datasets with different degrees of structure.
  • HE Jun, ZHANG Yunfei, ZHANG Dehai
    Computer Engineering. 2020, 46(11): 124-131. https://doi.org/10.19678/j.issn.1000-3428.0056429
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    To address the rule redundancy and logical conflicts of the sequential execution method of traditional rule chains applied to massive Data Cleaning(DC) task,this paper proposes an automatic combination and detection method for rule chains.A three-layer rule base including general,field-specific and customized layers is designed based on the context information.Then a Rule Chain Combination Model(RCCM) is established based on Petri Net(PN) to realize the automatic generation of rule chains,the detection of logical correctness and state accessibility,as well as the optimization of rule chains.The proposed method takes the DC application in the field of poverty alleviation in a certain area as an example.Experimental results on RCCM implementation show that the proposed method can significantly reduce generated error data and improves the quality and efficiency of DC.
  • MAO Yaqiong, TIAN Liqin, WANG Yan, MAO Yaping, WANG Zhigang
    Computer Engineering. 2020, 46(11): 132-138,147. https://doi.org/10.19678/j.issn.1000-3428.0056453
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    Existing outlier detection algorithms are generally time-consuming to deal with massive high-dimensional data streams.To address the problem,this paper proposes a Fast outlier detection algorithm in data stream with Local Density of Vector dot Product(FASTLDVP).It carries out incremental calculation only for the affected data points in the window,and keeps a small amount of intermediate results.Meanwhile,two optimization strategies and one pruning rule are designed to reduce the number of distance calculation times and the space-time overhead of the algorithm,so as to improve the detection efficiency.Theoretical analysis and experimental results show that this algorithm can effectively improve the detection efficiency of outliers in data stream while ensuring the detection accuracy,and can be extended to parallel environments for parallel acceleration.
  • WANG Mohan, ZHAI Junhai, QI Jiaxing
    Computer Engineering. 2020, 46(11): 139-147. https://doi.org/10.19678/j.issn.1000-3428.0055670
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    The Condensed Fuzzy K-Nearest Neighbor(CFKNN) algorithm is applicable only to small-sized and medium-sized data sets,and its mechanism of instance selection is static,so the algorithm can not adjust the threshold dynamically to select the optimal sample.To address the problems,this paper improves the CFKNN algorithm to make it applicable it to large-scale data environment,and on this basis proposes two kinds of large-scale condensed fuzzy K-nearest neighbor algorithms based on MapReduce and Spark.A dynamic mechanism is introduced into the setting of the instance selection threshold to make the selected instances more representative.Experiments are carried out on a big data platform with 7 data nodes.The experimental results show that compared with the CFKNN algorithm,the two proposed algorithms have better classification accuracy and acceleration ratio.Results of comparison between the two proposed algorithms show that MapReduce produces more intermediate files than Spark,yet Spark outperforms MapReduce in terms of running time and synchronization times.
  • Cyberspace Security
  • CAO Yongyi, JIN Weizheng, WU Jing, LUO Wei, ZHU Bo
    Computer Engineering. 2020, 46(11): 148-156. https://doi.org/10.19678/j.issn.1000-3428.0055831
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    In the Software Defined Network(SDN) architecture,most of the traditional Distributed Denial of Service(DDoS) attack detection mechanisms are based on the middle plug-ins or SDN controllers,which lacks the global network monitoring information and generates the high southbound interface communication overhead and detection delay.To address the problem,this paper proposes a DDoS attack detection and defense method based on cross plane cooperation in SDN architecture.The method uses the computing power of CPU of OpenFlow switch to offload part of the detection task from the control plane to the data plane,and then complete the whole detection task through the cooperation of the coarse-grained method of the data plane and the fine-grained method of the control plane.Based on the detection result,the controller formulates the defense strategy of the global scope of the network.Experimental results show that compared with the Support Vector Machine(SVM) method,the proposed method improves the detection efficiency and accuracy,decreases the detection delay and southbound interface communication overhead,and reduces the CPU load of the controller.
  • HU Bin, ZHOU Zhihong, YAO Lihong, LI Jianhua
    Computer Engineering. 2020, 46(11): 157-163. https://doi.org/10.19678/j.issn.1000-3428.0055588
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    The data sets for the detection of malicious traffic by the SSL/TLS protocol are single-sourced.Traditional detection methods take the quintuple feature of network traffic as the main feature for classification,which reduces the accuracy of malicious traffic detection in complex network environments.To address the problem,this paper proposes an improved method for encrypted malicious traffic detection.During data pre-processing,the encrypted malicious traffic is divided into two feature dimensions,packet payload and stream fingerprint,which are used to describe the distribution of traffic when the quintuple information is avoided.Also,the logistic regression model is used to realize the detection of encrypted malicious traffic.Experimental results show that,without relying on the five-tuple feature,the detection accuracy of the proposed method for malicious traffic encrypted by the SSL/TLS protocol in the complex network environment reaches 97.60%,which is approximately 36.05% higher than the traditional detection method based on quintuple feature and packet payload feature.
  • ZHANG Xiaolin, LIU Jiao, BI Hongjing, LI Jian, WANG Yongping
    Computer Engineering. 2020, 46(11): 164-173. https://doi.org/10.19678/j.issn.1000-3428.0056038
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    Existing privacy protection techniques are inefficient when applied to directed graphs of large-scale social networks,and publishing anonymous data does not meet the needs of community structure analysis.To address the problem,this paper proposes a K-In&Out-Degree Anonymity(KIODA) algorithm for large-scale social networks based on hierarchical community structure.The algorithm divides the community based on hierarchical community structure.The greedy algorithm is used to group K-in&out-degree sequences and make them anonymous,and the virtual nodes are added in parallel to achieve K-in&out-degree anonymity.Then information exchanges between nodes are implemented based on the GraphX platform.Virtual node pairs are selected based on the changes of the hierarchical community entropy,and are merged and deleted to reduce information loss.Experimental results show that the KIODA algorithm improves the efficiency of processing directed graphs of large-scale social networks,and ensures the availability of community structure analysis results in data publishing after the anonymity is realized.
  • GE Wenqi, YANG Qing, LIAO Junguo, HE Yuxuan
    Computer Engineering. 2020, 46(11): 174-180. https://doi.org/10.19678/j.issn.1000-3428.0056277
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    At present the rapidly growing malicious applications in Android systems have imposed significant threats to the security of mobile terminals,but the traditional detection systems fail to detect them quickly and effectively.To address the problem,this paper proposes a malware detection system which combines feature weighting with the deep learning algorithm using Bidirectional Long Short-Term Memory (Bi-LSTM) neural network.The static analysis method is used to extract different types of behavior features from malicious and normal applications.The feature weighting method is used to eliminate noise and irrelevant factors to construct feature vectors.The Bi-LSTM-based deep learning algorithm is used to optimize the behavior feature parameters.Then a classification model for malicious and normal applications is designed,and on this basis a detection system for malicious applications combining feature weighting and the deep learning algorithm is constructed.Experimental results show that compared with traditional detection systems such as Support Vector Machine(SVM) and RNN,the proposed system has higher precision and accuracy in malicious application detection.
  • CAO Suzhen, DU Xialing, YANG Xiaodong, LIU Xueyan, WANG Rui
    Computer Engineering. 2020, 46(11): 181-186,193. https://doi.org/10.19678/j.issn.1000-3428.0056432
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    Attribute-based searchable encryption technology can simultaneously meet the requirements of fine-grained access control and ciphertext data retrieval.This paper proposes a multi-keyword ciphertext retrieval scheme based on hybrid storage attributes using the anti-tampering and de-centralization features of blockchain.The scheme uses a public key encryption algorithm to encrypt the attribute key,ensuring that the attribute key assigned to the user can be securely transmitted on the public channel.The user revocation function is implemented by binding the user version number to the generation of the user attribute key to prevent the user from having unauthorized access to the data.Combining the technical advantages of blockchain,the ciphertext index and the ciphertext data are respectively stored in the blockchain and the cloud server to realize the verification of the correctness of the search results and the protection of data privacy.Simulation results based on difficult problems such as HDH,MDDH and CDH under the random prediction model prove the safety and efficiency of the scheme.
  • WU Mengli, CHEN Yuebin, WU Haifeng, LI Min, SUN Xiangsheng
    Computer Engineering. 2020, 46(11): 187-193. https://doi.org/10.19678/j.issn.1000-3428.0055930
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    To deal with the attacks of Spectrum Sensing Data Falsification(SSDF) in cognitive radio network,this paper proposes a Wiener Filter Detection(WFD) algorithm by using the Wiener filter based on the minimum Mean Square Error(MSE) to train the optimal weight and threshold for fusion decision.The algorithm uses the gradient algorithm to train the optimal weight,based on which the training data is weighted and fused,and the average of the fusion results is taken as the threshold.The weight obtained by training and the threshold are used to weight and fuse the data sent by each cognitive user to get the decision results.Simulation results show that compared with the traditional Equal Gain Combination(EGC) algorithm,the error probability of the WFD algorithm can be reduced by more than 20% under the same Signal-to-Noise Ratio(SNR).Also,the WFD algorithm has better robustness,and is less affected by the key parameters of SSDF attacks(including the proportion of malicious users,attack probability and relative attack intensity).
  • Mobile Internet and Communication Technology
  • CHEN Yuwan, JIA Xiangdong, JI Pengshan, Lü Yaping
    Computer Engineering. 2020, 46(11): 194-200. https://doi.org/10.19678/j.issn.1000-3428.0056001
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    To address the problem of describing the correlation between User Equipment(UE) and base station space in large-scale hot spot communication scenarios,this paper constructs a millimeter wave heterogeneous network model based on Poisson Cluster Process(PCP),and proposes a UE cluster classification scheme based on the nearest distance ratio of the Pico Base Station(PBS).According to the millimeter wave transmission model and path loss model,the expressions of the association probability of the UE cluster and the spectral efficiency of Down Links(DL) in millimeter wave heterogeneous networks are derived from the random geometry theory.The influence of transmission power,classification factor of UE cluster,maximum of PBS distribution on the association probability is analyzed,and the spectrum efficiency of users with Poisson cluster distribution and users with traditional Poisson point uniform distribution is compared.Simulation results show that compared with the traditional network model based on the Poisson Point Process(PPP),the proposed scheme significantly improves the DL spectral efficiency of the system model.
  • LI Xinying, HAO Hao, HUANG Haiyan
    Computer Engineering. 2020, 46(11): 201-206,213. https://doi.org/10.19678/j.issn.1000-3428.0056212
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    To address the low spectrum efficiency and low energy efficiency in cognitive network transmission,this paper proposes a slot allocation and transmission scheme based on simultaneous information and power transfer for cognitive relay networks.The transmission model for cognitive networks is established,in which the cognitive user and the primary user use spectrum transmission alternately.Then the relay node is introduced between the cognitive user’s source node and the destination node in order to collect and store the energy of the primary user’s transmitted signal when the primary user occupies the spectrum transmission.The energy is used to forward signals after the spectrum is released,and on this basis the expressions of the cognitive user’s throughput and energy efficiency under imperfect spectrum sensing reliability is deduced.Simulation results show that the transmission performance of the cognitive user of the proposed scheme increases with the spectrum sensing reliability,and the introduction of energy collection relay node for the cognitive user in low-power scenarios can effectively improve the throughput and energy efficiency of cognitive networks.
  • ZHAO Jihong, SUN Tianao, QU Hua, ZHANG Yin, ZHAI Fanni
    Computer Engineering. 2020, 46(11): 207-213. https://doi.org/10.19678/j.issn.1000-3428.0056549
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    In order to address the complex multi-controller deployment in large-scale Software Defined Network(SDN),this paper proposes an SDN controller deployment strategy based on the improved Louvain community detection algorithm.The strategy redefines the link weight according to the node similarity in the Louvain algorithm,and introduces the controller load difference to limit the number of nodes in each community in order to reduce the difference of the number of nodes between communities.At the same time,considering the influence of the three performance indicators including the propagation delay between the switch and the controller,the propagation delay between the controllers,and the reliability of the control link,a suitable location is selected to deploy the controller in each community.The simulation experiment results show that compared with the original Louvain algorithm and GABCC algorithm,the proposed algorithm can effectively reduce the propagation delay,balance the controller loads,and improve the reliability of the control link.
  • HAN Tingting, LIU Qiang, SUN Yantao, GUO Shoujiang
    Computer Engineering. 2020, 46(11): 214-222,230. https://doi.org/10.19678/j.issn.1000-3428.0055742
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    In order to improve the communication performance in the battlefield environment,this paper combines the EIGRP routing protocol with the tactical communication network according to the features of the heterogeneous tactical communication network.Based on the typical tactical organization system,this paper summarizes the architecture of the tactical communication network,and constructs the logic model of nodes in the hierarchical structure of the tactical communication network.On this basis,the basic principles of configuring the parameters of the EIGRP routing protocol in tactical communication network is proposed.Experiments in comparison are carried out by using the OPNET simulation tools.The results show that compared with the OSPF routing protocol,the EIGRP routing protocol provides reduced network recovery time and routing peak in tactical communication network.Also,it can produce routing results that meet different requirements by using different parameter configurations.
  • ZHAO Guofeng, LIN Huan, DUAN Jie, ZOU Yaqin, ZENG Shuai
    Computer Engineering. 2020, 46(11): 223-230. https://doi.org/10.19678/j.issn.1000-3428.0056463
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    The cache in Information-Centric Networking(ICN) can accelerate the transmission of data and reduce the delay of data response in Internet of Things(IoT).However,the existing ICN caching schemes do not consider the frequent data updates and users’ requests for data freshness,resulting in the low efficiency of caching.To solve the problem,this paper proposes an ICN-IoT caching scheme based on the freshness of IoT data.The timestamp is introduced to improve the time accuracy of user’s requests for data freshness.Then the data value in a future time is predicted based on the relevance of IoT data in order to meet the users’ requests.The content popularity and the time-based request probability are used to make a cache decision for the arrived content.Simulation results show that compared with NDN-PET,NDN-TTL and PCC caching schemes,the proposed scheme can effectively reduce the average delay for IoT users and improve the information accuracy rate.
  • ZHANG Jingyu, LIU Jingju, YE Chunming
    Computer Engineering. 2020, 46(11): 231-237,245. https://doi.org/10.19678/j.issn.1000-3428.0056406
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    To improve the sweep coverage with weighted targets and constrained return time,this paper proposes a region coverage algorithm,TLPS,based on target layering and path segmentation.The algorithm analyzes the location and weight information of all targets and translate these targets into points.According to the point information,the algorithm calculates the location of the sink and hierarchically extracts weighted targets.For the point target set between the same levels,the TSP path based on the greedy policy is designed.On this basis,a loop segmentation strategy for weight nodes is designed,and TSP paths are re-segmented to obtain the final scan paths.Experimental results show that by adding a small number of sensor nodes,the TLPS algorithm has a shorter average scan period,higher target coverage efficiency and higher path effectiveness rate than tcwtp,OSweep and other algorithms.
  • SHI You, LIN Lin, WU Yan
    Computer Engineering. 2020, 46(11): 238-245. https://doi.org/10.19678/j.issn.1000-3428.0055963
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    When multiple mobile nodes in Wireless Sensor Networks(WSNs)collect,receive and send data,the irregularity of data sources will cause nodes to receive a large amount of data instantaneously,resulting in serious network congestion.However,the existing network congestion control methods have low calculation accuracy.To solve the problem,this paper proposes a fuzzy Proportional-Integral-Differential(PID) congestion control model,CFPID,based on the Cuckoo Search(CS) algorithm.The PID controller is introduced into WSNs.Then the fuzzy control algorithm is used to adjust and optimize PID parameters to improve calculation accuracy,and the CS algorithm is used to search and optimize the quantization factor and parameter increment of fuzzy PID control,so as to realize the precise control of message queue length in nodes.Experimental results show that compared with the traditional control methods such as PID and IBLUE,the proposed method has stronger control ability on queue length and packet loss rate,and the network throughput is increased by 4%~8%.
  • Graphics and Image Processing
  • HE Zhicheng, WANG Zhenxing
    Computer Engineering. 2020, 46(11): 246-254. https://doi.org/10.19678/j.issn.1000-3428.0056446
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    The machine-vision-based automatic detection methods for the welding spots of body in white provides an effective way to control the quality of the welding.However,due to the influence of light pollution,the machine vision system of the automatic detection equipment often fail to locate the welding spots accurately.To address the interference from environment and the weak robustness of the traditional methods,this paper proposes a method for welding spot detection based on deep learning.The method introduces the convolutional structure of MobileNetv2 to replace the convolutional layer of YOLOv2,and draws on the fine-grained feature of YOLOv2 to solve the excessive number of parameters of the traditional YOLOv2 model.Then GIoU loss is used to improve the loss function of the model.Finally,the K-means clustering algorithm is used to obtain a suitable anchor for the welding spot dataset,and an efficient and reliable welding spot detection model FGM_YOLO for light body in white is obtained.Results of testing and comparison on the test set of welding spots of body in white show that the AP of this model is 2.47% higher than that of YOLOv2.The number of its parameters is about 1/16 of that of the original model,and its detection speed is two times that of the original model,which proves a significant increase in the detection accuracy and efficiency.
  • XU Xiaoyu, ZHAO Longzhang, CHENG Xiaoyue, HE Zhichao
    Computer Engineering. 2020, 46(11): 255-260,266. https://doi.org/10.19678/j.issn.1000-3428.0055858
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    Traditional Convolutional Neural Network(CNN) are prone to overfitting,and have performance in generalizing random data.To address the problem,this paper designs a CNN based on Fisher discriminant criterion and GRV module,called FDCNN.The method uses a loss function based on improved Fisher discriminant criterion to train the model,and maps the sample data of human faces to low-dimensional space.So the dispersion of the same type of faces in the mapping space is minimized while the dispersion of different types of faces is maximized to achieve the optimal face classification performance.In addition,this paper combines the advantages of GoogleNet,ResNet,VGGNet network structures to design a new GRV module,which improves the representation ability of the CNN and simplifies the network model.Experimental results show that when the number of training samples is 840,the recognition rate of the proposed FDCNN model on the CBCL dataset reaches 93.4%,which outperforms the traditional CNN model,fully connected neural network model based on improved Fisher discriminant criterion,and other network models.Also,the FDCNN model has better generalization ability than the above models.
  • YE Qinghao, TU Daijian, BI Qi, QIN Feiwei, GE Ruiquan, BAI Jing
    Computer Engineering. 2020, 46(11): 261-266. https://doi.org/10.19678/j.issn.1000-3428.0056527
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    When applied to security checks,common deep learning methods have to address the high false alarm rate caused by the unknown size,shape and location of hidden objects as well as unbalanced sample categories.To deal with the problem,this paper proposes a deep convolutional neural network model based on multiple view architecture.The model uses convolutional neural networks with residual connections to extract features.Then a Long Short Term Memory(LSTM) attention model based on dense connections is used to simulate the process of human observations from multiple perspectives to enhance the expression of threat-related information.At the same time,the network is optimized based on the focus loss function to form an end-to-end framework.The experimental results on HD-AIT millimeter-wave scaned human body threating dataset show that the proposed model increases the accuracy to 0.997 and recall rate to 0.999 compared with other baseline models.
  • JIA Rong, WANG Feng, YUAN Hongwu, TUO Haonan, JIANG Zhaozhen, WU Yunzhi
    Computer Engineering. 2020, 46(11): 267-272,278. https://doi.org/10.19678/j.issn.1000-3428.0055925
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    The ultraviolet polarization image of latent fingerprint is fused of ultraviolet intensity image and polarization parameter image,which can realize accurate detection and recognition of latent fingerprint.However,it is impossible to select the optimal polarization parameter to represent the target characteristics.Based on the existing fusion algorithms for the polarization image,this paper proposes a fuzzy adaptive fusion algorithm for the ultraviolet polarization image.It analyzes multiple polarization parameter images extracted from the ultraviolet polarization image,and uses fuzzy integral to select the optimal polarization parameter image adaptively.Then the ultraviolet intensity image and the optimal polarization parameter image are decomposed into high and low frequency coefficients by using discrete Stationary Wavelet Transform(SWT).Also,the high frequency coefficients are fused based on the maximum rule,and the low frequency coefficients are fused based on the sparse representation rules.On this basis,the inverse discrete SWT is used to obtain the fused image.Experimental results show that compared with LP,PCA and other fusion algorithms,the proposed algorithm can better retain the features of intensity images and high frequency information of polarization parameters in the fused images.It improves the target contrast and enhances the target detail features,and has strong adaptability to different materials of latent fingerprints.
  • WANG Wenbing, LIU Fenlin, GONG Daofu, LIU Shengli
    Computer Engineering. 2020, 46(11): 273-278. https://doi.org/10.19678/j.issn.1000-3428.0056668
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    The method of using singular value robustness to embed grayscale watermarks can embed a large number of watermarks,and has high invisibility and robustness,but its practicability is affected by the false alarm problem.To this end,a new reliable watermarking scheme is proposed by MAKBOL et al,but it does not completely solve the false alarm problem.Therefore,this paper proposes an attack approach by analyzing the embedding and extracting process of MAKBOL’s scheme and using the strong correlation between its side information and watermark.The approach forges side information to extract the false watermark from the extraction process of MAKBOL’s scheme.Experimental results show that this attack approach not only works for carriers without embedded watermarks,but also works for carriers with embedded watermarks.
  • Development Research and Engineering Application
  • XU Guozheng, LIAO Chencong, CHEN Jinjian, DONG Bin, ZHOU Yue
    Computer Engineering. 2020, 46(11): 279-285. https://doi.org/10.19678/j.issn.1000-3428.0056061
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    Existing detection methods for apparent crack of concrete structure are inaccurate and low-precision,losing much detail information.To address the problem,this paper proposes an apparent crack detection method for concrete based on the HU-ResNet Convolutional Neural Networks(CNN).Based on improved U-Net,the HU-ResNet model is established using the ResNet34 residual network trained by ImageNet as the encoder to retain crack details and accelerate network convergence.The scSE attention mechanism module is also introduced to recalibrate the output characteristics of the encoding block and decoding block in space and channel.At the same time,the output feature maps of each stage of the decoder are fused by the hypercolumn module to obtain more accurate semantic information and location of crack images,and the precision of which is further improved by using the combined loss function.Experimental results show that the pixel accuracy,Intersection-over-Union and F1 value of the proposed model reach 0.990 4,0.693 3 and 0.816 6 respectively,which are better than that of Canny, region growing and other traditional digital image models and FCN8s,U-Net,U-ResNet and other deep learning models,and the proposed model has more accurate crack detection results.
  • DING Yuan, HAO Minglei, XING Hongyan, ZENG Xiangneng
    Computer Engineering. 2020, 46(11): 286-292,300. https://doi.org/10.19678/j.issn.1000-3428.0055736
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    The inversion of the ocean surface wind field with single-polarized Synthetic Aperture Radar(SAR) data has a complex operational model.The use of cross polarization SAR images to invert the ocean surface wind has become a hot topic in research.Taking the waters of the Pacific and the Atlantic as the research objects,this paper uses the polarized SAR image data from the GF-3 satellite,which is the C-band SAR satellite launched independently by China,to analyze the relationship between cross polarization backscattering intensity and ocean surface wind speed,relative wind direction and radar incident angle.A multiple linear regression model and a BP neural network model are established,and their results are verified by using wind field reanalysis data of ECMWF.Experimental results show that the established regression model demonstrates the linear correlation between the cross polarization backscattering intensity and the wind speed and incident angle.Compared with the inversion of sea surface wind field based on co-polarization SAR,this inversion process does not depend on the input of the external wind direction,and simplifies the wind speed inversion model.The R value in the fit of training sample set of the BP neural network model exceeds 70%,which means the model effectively predicts the cross polarization wind speed.
  • WANG Ruoyu, CHEN Yongquan
    Computer Engineering. 2020, 46(11): 293-300. https://doi.org/10.19678/j.issn.1000-3428.0055910
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    Among the existing algorithms for TSP solution,the heuristic algorithm based on Iterated Local Search(ILS) performs the best,holding the world record on most of the public instances.The method for solution construction has a significant influence on the performance of ILS,and thus should be carefully designed.This paper proposes four different methods for solution construction,including a baseline algorithm that uses only static information such as the distances between cities to construct the initial solution,and three reinforcement-learning-based algorithms that attempt to utilize reinforcement learning to dig useful information from the historic information collected during the search for the construction of initial solutions.Experimental results on 25 public instances show that the reinforcement-learning-based methods using historic information can significantly improve the quality of the constructed solution as well as the performance of ILS.
  • ZHENG Wenxiu, ZHAO Junyi, WEN Xinyi, YAO Yindi
    Computer Engineering. 2020, 46(11): 301-305,314. https://doi.org/10.19678/j.issn.1000-3428.0056278
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    The Mel-Frequency Cepstral Coefficient(MFCC) speech features cannot effectively reflect the effective information between consecutive frames.To address the problem,this paper uses deep neural network to extract bottleneck features with long-term correlation and compactness of speech,and on this basis proposes a compound feature construction method that combines the neural network bottleneck features and the MFCC feature,so as to improve the speech characterization and modeling capabilities.The MFCC feature is extracted from the speech data as the input,and then concatenated with the BN feature to obtain a new compound feature.On this basis the acoustic modeling of Mixture Model-Hidden Markov Model(GMM-HMM) is implemented.Experimental results on the TIMIT database show that compared with the methods based on the single bottleneck feature and deep neural network posterior feature,the proposed method can significantly increases the recognition rate.
  • XU Guowei, CHEN Jian, CHENG Yi
    Computer Engineering. 2020, 46(11): 306-314. https://doi.org/10.19678/j.issn.1000-3428.0055955
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    Modern radar clutter simulation needs to use clutter data for real-time analysis and processing of the echo signal.However,the traditional Spherically Invariant Random Process (SIRP) method for clutter data generation is time-consuming.By improving the SIRP method,this paper proposes a method to improve the real-time performance of clutter generation based on Graphic Processing Unit (GPU) parallel computing.In the Compute Unified Device Architecture (CUDA),the multi-task series-parallel analysis is carried out for the correlation coherent K-distribution clutter algorithm.In addition,the cuBLAS library is used to optimize the fine-grained convolution calculation,and the OpenMP + CUDA multi-task scheduling mechanism is used to improve the coarse-grained task parallel calculation in order to improve the CPU-GPU utilization and reduce the data waiting time.Experimental results show that compared with the traditional GPU method,the proposed method increases the calculation speed by 61%,and the real-time performance of clutter data generation is effectively improved.In addition,the acceleration effect significantly grows with the volume of clutter data.
  • YU Liangjun, GAN Shengfeng, FAN Zhengwei
    Computer Engineering. 2020, 46(11): 315-320. https://doi.org/10.19678/j.issn.1000-3428.0056498
    Abstract ( ) Download PDF ( ) HTML ( )   Knowledge map   Save
    In the field of data mining and machine learning, classification is a key problem to which the Bayesian network model is frequently applied due to its simplicity and high efficiency.As a classical Bayesian network model for semi-supervised learning,One-Dependence Estimator(ODE) has been widely concerned by researchers.However,the existing ODE model classifiers do not consider the varying contribution of different attribute nodes acting as root nodes to the classification process.Therefore,this paper combines ODE model classifier with the attribute value weighting method,and on this basis proposes the MI-ODE algorithm.The algorithm adopts Mutual Information(MI) to measure the dependence between attribute values and class variables of the attribute root node,which is used as the weight of the ODE model.Then weighted average is implemented for the attribute values of the ODE classifier model.The MI-ODE algorithm is tested on 36 standard data sets for real-world classification problems,and results show that compared with NB algorithm,AODE algorithm and TAN algorithm,the proposed algorithm has better classification performance.