Computer Science

Survey of Influence Analysis of Evolutionary Network Based on Big Data

HE Qiang, YIN Zhen-yu, HUANG Min, WANG Xing-wei, WANG Yuan-tian, CUI Shuo, ZHAO Yong

Computer Science. 2022, 49 (8): 1-11. doi:10.11896/jsjkx.210700240

Abstract

PDF(2110KB) ( 958 )

References | Related Articles | Metrics

One of the most important technologies in modern information and service industry is social influence analysis.More and more researchers in social networks focus on social influence.Real social networks are evolving rather than static.The proposal of evolutionary network also brings new challenges and opportunities.At the same time,the massive social information in the evolutionary network also provides strong support for the rapid development of big data analysis technology.In this paper,evolutionary network and influence maximization are discussed.It also discusses the diffusion model of social influence analysis and the influence analysis method based on social network big data.At the same time,some widely used influence algorithms are further sorted out.In addition,this paper also discusses the relationship between big data,evolutionary networks,and social influence maximization.This paper aims to help other researchers to better understand the existing work and provide new ideas for the influence analysis of social networks through the influence analysis of large-scale social networks.

Survey of Multi-label Classification Based on Supervised and Semi-supervised Learning

WU Hong-xin, HAN Meng, CHEN Zhi-qiang, ZHANG Xi-long, LI Mu-hang

Computer Science. 2022, 49 (8): 12-25. doi:10.11896/jsjkx.210700111

Abstract

PDF(3598KB) ( 1566 )

References | Related Articles | Metrics

Most of the traditional multi-label classification algorithms use supervised learning,but in real life,there are many unlabeled data.Manual tagging of all required data is costly.Semi-supervised learning algorithms can work with a large amount of unlabeled data and labeled data,so they have received more attention from people.For the first time,multi-label classification algorithms are explained from the perspective of supervised learning and semi-supervised learning,and application fields of multi-label classification algorithms are comprehensively summarized.Among them,supervised learning algorithms of label non-correlation and label correlation are described in terms of decision trees,Bayesian,support vector machines,neural networks,and ensemble,semi-supervised learning algorithms are summarized from the perspectives of batch and online learning.The real-world application areas are introduced from the perspectives of image classification,text classification and other fields.Secondly,this paper briefly introduces evaluation metrics of multi-label.Finally,research directions of complex concept drift under semi-supervised learning,feature selection,complex correlation of labels and class imbalance are given.

Accelerating Persistent Memory-based Indices Based on Hotspot Data

LIU Gao-cong, LUO Yong-ping, JIN Pei-quan

Computer Science. 2022, 49 (8): 26-32. doi:10.11896/jsjkx.210700176

Abstract

PDF(2090KB) ( 974 )

References | Related Articles | Metrics

Non-volatile memory(NVM),also known as persistent memory(PM),has the characteristics of bit-based addressing,durability,high storage density and low latency.Although the latency of NVM is much smaller than that of solid-state drives,it is greater than that of DRAM.In addition,NVM has shortcomings such as unbalanced reading and writing as well as short writing life.Therefore,currently NVM cannot completely replace DRAM.A more reasonable method is using NVM to build a hybrid memory architecture based on DRAM+NVM.Based on the observation that many data accesses in database applications are skewed,this paper focuses on the hybrid memory architecture composed of NVM and DRAM and proposes a hotspot data-based speedup method for persistent memory indices.Particularly,we utilize the low latency of DRAM and the durability and high sto-rage density of NVM,and propose to add a DRAM-based hotspot-data cache for persistent memory indices.Then,we present a query-adaptive indexing method that can automatically adjust the cache according to the change of hotspot data.We apply the proposed method to several persistent memory indices,including wBtree,FPTree and Fast&Fair,and conduct comparative experiments.The results show that when the number of hotspot data visits accounts for 80% of the total visits,the proposed method can accelerate the query performance of the three indices by 52%,33% and 37%,respectively.

Spatio-Temporal Attention-based Kriging for Land Deformation Data Interpolation

LI Rong-fan, ZHONG Ting, WU Jin, ZHOU Fan, KUANG Ping

Computer Science. 2022, 49 (8): 33-39. doi:10.11896/jsjkx.210600161

Abstract

PDF(2740KB) ( 823 )

References | Related Articles | Metrics

Landslide is one of the most common geological hazards,it causes significant damage to people’s life and property everyyear.In order to prevent and control landslides,it is necessary to monitor the land surface extensively.However,insurmountable difficulties such as severe climate and high monitoring cost impede the collection of land surface data,resulting in incomplete local data,unbalanced data sampling and dynamic changes of monitoring points,which hinder the prevention and control research of landslide and put forward new demand to the data collection and analysis.Existing methods try to handle incomplete data from spatial perspective,which,however,ignore temporal dependencies that are important for data interpolation.To solve the above problems,the incomplete INSAR data filling is studied,the spatio-temporal dependence is modeled by using the spatio-temporal mask matrix,the multi-level spatial relationship is comprehensively studied by using multi-head attention,and a novel Kriging interpolation method using spatio-temporal attention is proposed on the basis of Kriging.It realizes the deep understanding of complex temporal and spatial features.Interpolation experiments on real-world INSAR datasets show that the proposed model is capable to learn sophisticated spatial and temporal features effectively,and achieves better performance than the state-of-the-art methods in three different data interpolation scenarios.

Multi-time Scale Spatial-Temporal Graph Neural Network for Traffic Flow Prediction

WANG Ming, PENG Jian, HUANG Fei-hu

Computer Science. 2022, 49 (8): 40-48. doi:10.11896/jsjkx.220100188

Abstract

PDF(3049KB) ( 1136 )

References | Related Articles | Metrics

Traffic flow prediction plays a key role in intelligent transportation system.However,due to its complex spatial-temporal dependence and its uncertainty,research becomes extremely challenging.Some existing methods mainly use a single time series and input it into a recurrent neural network to capture temporal dependency.Moreover,most models only simply stack temporal modules and spatial modules,resulting in ineffective feature fusion.To address these issues,this paper proposes a multi-time scale spatial-temporal graph neural network model.The model divides the sequence data into three time-scale sequences,then puts sequence data into the ST-Blocks to extract the spatial-temporal features of data,and finally makes the prediction.In ST-Block,graph convolutional network and variant Transformer are used to capture spatial dependency and temporal dependency respectively,and the output feature of the two sub-modules are fused through a gate mechanism.A large number of experiments are conducted on two real data sets in this paper,including short-term and long-term prediction,and the results show the excellent performance of MTSTGNN model on the task of traffic flow prediction.

Query Performance Prediction Based on Physical Operation-level Models

WANG Run-an, ZOU Zhao-nian

Computer Science. 2022, 49 (8): 49-55. doi:10.11896/jsjkx.210700074

Abstract

PDF(3008KB) ( 496 )

References | Related Articles | Metrics

Query performance prediction (QPP) is an important issue in database systems.When there are concurrent transactions in a database system,the existing methods fail to establish an accurate model without changing query performance.In this paper,a new method is proposed to solve the QPP problem.The proposed method builds unit prediction models for various physical operations in the query and combines the unit models into a complete QPP model according to the query plan.It can describe the concurrency state of the database system by taking the statistical information as features.The proposed method only needs to use the basic means provided by the DBMS to obtain the database statistics required to build the model,without changing the DBMS or affecting the execution of the original workloads on the database system.We evaluate our techniques on various workloads including OLTP and OLAP.Experiments show that the proposed method outperforms the state-of-art QPP methods regardless of different query plans or different concurrency.

RIIM:Real-Time Imputation Based on Individual Models

LI Xia, MA Qian, BAI Mei, WANG Xi-te, LI Guan-yu, NING Bo

Computer Science. 2022, 49 (8): 56-63. doi:10.11896/jsjkx.210600180

Abstract

PDF(3257KB) ( 444 )

References | Related Articles | Metrics

With the enrichment of data sources,data can be obtained easily but with low quality,resulting that the MVs are ubi-quitous and hard to avoid.Consequently,MV imputation has become one of the classical problems in the field of data quality mana-gement.However,most existing MV imputation approaches are proposed for static data,which cannot handle dynamic data streams arriving at high-speed.Moreover,they do not consider data sparsity and heterogeneity simultaneously.Therefore,a novel MV imputation approach,real-time imputation based on individual models (RIIM) is proposed.In RIIM,the MVs are effectively filled by combining the basic ideas of neighbors-based imputation and regression-based imputation with consideration of sparsity and heterogeneity of data.For the dynamic and real time of data streams,the MV imputation model is updated incrementally.Moreover,an adaptive and periodic updating strategy for optimal neighbors search is proposed to solve the problem of high time cost and hard to determine the number of neighbors.Finally,the effectiveness of the proposed RIIM is evaluated based on extensive experiments over real-world datasets.

Hierarchical Granulation Recommendation Method Based on Knowledge Graph

QIN Qi-qi, ZHANG Yue-qin, WANG Run-ze, ZHANG Ze-hua

Computer Science. 2022, 49 (8): 64-69. doi:10.11896/jsjkx.210600111

Abstract

PDF(2724KB) ( 735 )

References | Related Articles | Metrics

The recommendation system based on graph neural network is the current research hotspot of data mining applications.The recommendation performance can be improved by combining the graph neural network on the heterogeneous information network(HIN).However,the existing HIN-based recommendation methods often have problems that cannot effectively explain the results of high-level modeling,and manual design of meta-paths requires knowledge of related domains.Therefore,this paper combines the idea of hierarchical granulation andproposes a heterogeneous recommendation method(HKR) based on knowledge graphs.The local context and non-local context are hierarchically granulated,and the coarse-grained representation of user characteristics is learned separately.Then based on the gating mechanism, combining local and non-local attribute node embedding,learning the potential features between users and items,and finally fusing fine-grained features for recommendation.The real experimental results show that the performance of the proposed method is better than the current graph neural network recommendation method based on knowledge graph in many aspects.

Cross-domain Recommendation Algorithm Based on Self-attention Mechanism and Transfer Learning

FANG Yi-qiu, ZHANG Zhen-kun, GE Jun-wei

Computer Science. 2022, 49 (8): 70-77. doi:10.11896/jsjkx.210600011

Abstract

PDF(2520KB) ( 698 )

References | Related Articles | Metrics

Traditional single-domain recommendation algorithm is limited by the sparse relationship between users and items,and there is a problem of user/item cold start,and only models the item ratings by users,ignoring the information contained in the review text.The cross-domain recommendation algorithm based on review text extracts user/item review information in the auxiliary domain to alleviate data sparseness in the target domain and improve the accuracy of recommendation.This paper proposes a cross-domain recommendation algorithm SAMTL that combines self-attention mechanism and transfer learning.Different from existing algorithms,SAMTL fully integrates the knowledge of the target domain and auxiliary domains.Firstly,the self-attention mechanism is introduced to model user’s preference information.Then,by the cross-mapping cross-domain transmission network,the recommendation accuracy of another domain is improved with the help of information in one domain.Finally,the information of the two domains is integrated in the knowledge fusion and scoring prediction module to perform scoring prediction.Experiments on Amazon data set show that,compared with the existing cross-domain recommendation model,SAMTL has higher MAE and MSE values,and MAE increases by 8.4%,13.2% and 19.4% on three different cross-domain data sets,MSE increases by 6.3%,7.8% and 5.6% respectively.A number of experiments verify the effectiveness of self-attention mechanism and transfer learning,as well as the advantages in alleviating data sparsity and user cold start problems.

Analysis Method of APP User Behavior Based on Convolutional Neural Network

CHEN Yong-quan, JIANG Ying

Computer Science. 2022, 49 (8): 78-85. doi:10.11896/jsjkx.210700121

Abstract

PDF(2029KB) ( 932 )

References | Related Articles | Metrics

With the rapid development of mobile Internet,smart terminal has become an indispensable part of people’s daily life and work.In the process of using smart terminal,a large number of APP operation process records will be generated.By analyzing the user’s APP operation process records,the user’s behaviors in the operation process and the user’s behavior pattern can be obtained,which can help developers maintain and improve the APP software.Existing user behavior analysis is biased towards operation analysis and thebehaviors extraction for user’s operation is lacked.An APP user behaviors analysis method based on convolution neural network is proposed.At first,the APP operations are analyzed,and the user operations in the original APP operation record information are extracted.Then the correlation between APP operation and user’s behaviors is mined,and the similarity matrix between APP operations and APP user’s behaviors is constructed.Finally,the behaviors of users will be extracted.Experiments show that this method can extract and identify the user’s behaviors in the records of APP operationprocess effectively,which will be helpful to explore the deep meaning of user’s behaviors.

Unsupervised Multi-view Feature Selection Based on Similarity Matrix Learning and Matrix Alignment

LI Bin, WAN Yuan

Computer Science. 2022, 49 (8): 86-96. doi:10.11896/jsjkx.210700124

Abstract

PDF(4512KB) ( 642 )

References | Related Articles | Metrics

Multi-view feature selection improves the efficiency of classification,clustering and other learning tasks by fusing information from multiple views to obtain representative feature subsets.However,the features of different views that describe objects are complex and interrelated.Simply searching subset of features from original space partly solves the problem of dimension,but it barely obtains the latent structural information and association information among features.Besides,using fixed similarity matrix and projection matrix is prone to lose the correlation between different views.To solve these problems,an unsupervised multi-view feature selection algorithm based on similarity matrix learning and matrix alignment(SMLMA)is proposed.Firstly,the similarity matrix based on all views is constructed,and the consistent similarity matrix and projection matrix are obtained by mani-fold learning,to explore and reserve the structural information of data to the greatest extent.Then,the matrix alignment method is used to maximize the correlation between the similarity matrix and the kernel matrix,for the purpose of using the correlation between different views and reducing the information redundancy of feature subset.Finally,the Armijo searching method is introduced to obtain the convergence result quickly.Experimental results on four datasets(Caltech-7,NUS-WIDE-OBJ,Toy Animal and MSRC-v1)show that,compared with single view feature selection and some multi-view feature selection methods,the accuracy of SMLMA is averagely improved by about 7.54%.The proposed algorithm well retains the structural information of data and the correlation between multi-view features,and captures more high-quality features.

Strongly Connected Components Mining Algorithm Based on k-step Search of Vertex Granule and Rough Set Theory

CHENG Fu-hao, XU Tai-hua, CHEN Jian-jun, SONG Jing-jing, YANG Xi-bei

Computer Science. 2022, 49 (8): 97-107. doi:10.11896/jsjkx.210700202

Abstract

PDF(3147KB) ( 459 )

References | Related Articles | Metrics

Strong connected components (SCCs) mining is one of the classic problems in graph theory.It has practical requirements to design a serial SCCs mining algorithm with high efficiency.GRSCC algorithm can use SUB-RSCC function to discover SCCs of simple digraphs.SUB-RSCC function is formed by two operators of rough set theory (RST),k-step upper approximation set and k-step R-related,which are the main contributors to time consumption.Then the invocation times of SUB-RSCC decide the efficiency of GRSCC algorithm.Based on the SCCs correlations among vertices,GRSCC algorithm introduces granulation strategy to reduce the invocation times of SUB-RSCC function,then improve the mining efficiency.Two new SCCs correlations are found by analysis of SCCs in the framework of RST,then a new vertex granulation strategy is designed to granulate the vertex set of target digraphs.In order to reduce the invocation times of SUB-RSCC function to a greater extent,a method called k-step search of vertex granule is proposed.Finally,combining with GRSCC algorithm,an algorithm called KGRSCC for mining SCCs based on k-step search of vertex granule and RST is proposed.Experimental results show that,compared with RSCC,GRSCC and Tarjan algorithms,the proposed KGRSCC algorithm has better performance.

Mixed Attribute Feature Detection Method of Internet of Vehicles Big Datain Multi-source Heterogeneous Environment

CHEN Jing, WU Ling-ling

Computer Science. 2022, 49 (8): 108-112. doi:10.11896/jsjkx.220300273

Abstract

PDF(1757KB) ( 552 )

References | Related Articles | Metrics

Current feature detection methods for big data of Internet of vehicles ignore the data attribute weight,resulting in low efficiency and fail to provide efficient services in vehicle operation.Therefore,a hybrid attribute feature detection method for big data of Internet of vehicles in multi-source heterogeneous environment is proposed.Middleware is used to build an integration model to integrate multi-source heterogeneous data of the Internet of vehicles,and standardization and attribute reduction of integrated data are completed.With pre-processed data as input,attribute features are extracted by weighted principal component analysis,and feature clustering is realized by clustering method to complete the feature detection of mixed attribute of Internet of vehicles big data.Experimental results show that compared with existing methods,the sensitivity index of the proposed method is higher and the time complexity is lower,which indicates that the proposed feature detection method is more efficient and can more accurately complete the feature extraction task of mixed attributes of the Internet of vehicles big data.

Deep Hash Retrieval Algorithm for Medical Images Based on Attention Mechanism

ZHU Cheng-zhang, HUANG Jia-er, XIAO Ya-long, WANG Han, ZOU Bei-ji

Computer Science. 2022, 49 (8): 113-119. doi:10.11896/jsjkx.210700153

Abstract

PDF(1923KB) ( 582 )

References | Related Articles | Metrics

A medical image retrieval method combining attention mechanism is proposed for a series of problems such as poor retrieval performance,low accuracy and lack of interpretability in current medical image retrieval.Based on deep convolutional neural networks and taking Bayesian models as the framework,the proposed algorithm introduces an attention mechanism module guided by semantic features.Local feature descriptors containing semantic information are generated under the guidance of the classification network.Both global features and local features rich in semantic information are used as inputs to the hash network,which enhances the feature representation capability of hash coding by guiding the hash network to pay attention to important feature regions from both global and local perspectives.And the weighted likelihood estimation function is introduced to solve the problem of the unbalanced number of positive and negative sample pairs.MAP and NDCG are used as evaluation metrics,and the ChestX-ray14 dataset is selected for experiments.The proposed algorithm is compared with the current commonly used deep ha-shing methods.Experiment results show that the MAP and NDCG values are much better than the existing deep hashing methods at different code levels of hash coding,which proves the effectiveness of the proposed algorithm.

Re-parameterized Multi-scale Fusion Network for Efficient Extreme Low-light Raw Denoising

WEI Kai-xuan, FU Ying

Computer Science. 2022, 49 (8): 120-126. doi:10.11896/jsjkx.220200179

Abstract

PDF(2900KB) ( 647 )

References | Related Articles | Metrics

Practical low-light denoising/enhancement solutions often require fast computation,high memory efficiency,and can achieve visually high-quality restoration results.Most existing methods aim to restore quality but compromise on speed and memory requirements,which limits their usefulness to a large extent.This paper proposes a new deep denoising architecture,a re-parameterized multi-scale fusion network for extreme low-light raw denoising,which greatly improves the inference speed without losing high-quality denoising performance.Specifically,image features are extracted in multi-scale space,and a lightweight spatial-channel parallel attention module is used to focus on core features within space and channel dynamically and adaptively.The representation ability of the model is further enriched by re-parameterized convolutional unit without increasing computational cost at inference.The proposed model can restore UHD 4K resolution images within about 1s on a CPU(e.g.,Intel i7-7700K) and run at 24 fps on a GPU(e.g.,NVIDIA GTX 1080Ti),which is almost four times faster than existing advanced methods(e.g.,UNet) while still maintaining competitive restoration quality.

Reversible Hidden Algorithm for Remote Sensing Images Based on Hachimoji DNA and QR Decomposition

WANG Kun-shu, ZHANG Ze-hui, GAO Tie-gang

Computer Science. 2022, 49 (8): 127-135. doi:10.11896/jsjkx.210700216

Abstract

PDF(5000KB) ( 497 )

References | Related Articles | Metrics

In recent years,with the rapid development of cloud computing and artificial intelligence,digital images are more and more widely used in multimedia,medical and military fields.Among them,remote sensing technology can detect and process remote targets through the theory of electromagnetic wave,which makes the transmission security of remote sensing images increasingly significant.In order to solve the problems of digital image information security and privacy protection,this paper proposes a reversible remote sensing image hiding algorithm based on novel Hachimoji deoxyribonucleic acid (DNA) and QR decomposition.Firstly,the system parameters and initial values of the coupled map lattice(CML) are updated by using the information entropy of remote sensing image and host image,and the generated chaotic sequences are used as the one-time key stream of the encryption process,which increases resistance to known/selected plain-text attacks.Then,Hachimoji DNA technology is used to encode the remote sensing image with 8-bit base,and key streams are used to perform XOR operation and cyclic shift operation on the image matrix.Finally,the host image is decomposed into blocks,and the encrypted remote sensing image is embedded into the host block in the form of DNA bases.In particular,the host image still visually meaningful after embedding remote sensing information,and the embedded image information can also be extracted from it without loss.Simulation results show that the proposed algorithm has good embedding effect,high security and strong robustness.

Incremental Object Detection Method Based on Border Distance Measurement

LIU Dong-mei, XU Yang, WU Ze-bin, LIU Qian, SONG Bin, WEI Zhi-hui

Computer Science. 2022, 49 (8): 136-142. doi:10.11896/jsjkx.220100132

Abstract

PDF(2816KB) ( 625 )

References | Related Articles | Metrics

Incremental learning has achieved good results in image classification,but it is challenging to apply incremental learning to multi-class object detection.Object detection is more complex than image classification,which combines classification and border regression.At present,the most advanced incremental object detectors adopt the external fixed region suggestion method based on knowledge distillation,which consumes a lot of time and cost.For single-stage detectors,due to the lack of annotation and region advice information for the old class,old objects are usually identified by the detector as the background,resulting in catastrophic forgetting.In this paper,a label selection algorithm based on border distance metric is proposed.It uses the detection results of the old model and the existing dataset labels to select and merge by measuring the coincidence of the bounding boxes,making up for the lack of annotations of the old objects in the new dataset and alleviating catastrophic forgetting.In addition,a module that combines the attention module with the residual module is designed to extract discriminative features at different depths in feature extraction network,to further improve the detection accuracy of model.The proposed method is implemented in the single-stage detection framework,and the effectiveness of the method is verified on PASCAL VOC dataset.Compared with the best model at present,the average accuracy value of the old object and all objects improves by 2.8% and 2.1%,respectively.The pseudo-labels obtained by the proposed method greatly alleviate the forgetting problem,and the attention residual module improves the detection accuracy of the model.

Spatial Multi-feature Segmentation of 3D Lidar Point Cloud

YANG Wen-kun, YUAN Xiao-pei, CHEN Xiao-feng, GUO Rui

Computer Science. 2022, 49 (8): 143-149. doi:10.11896/jsjkx.210300275

Abstract

PDF(2871KB) ( 493 )

References | Related Articles | Metrics

Multi-layer solid-state lidar has become an important tool for environment perception of unmanned platform,and has been widely used in vehicle-mounted environment modeling.Due to the low resolution of lidar,the sensitivity of environmental noise,and the complexity of the scene,the fast and effective segmentation of the scene becomes a key problem in the real-time environment modeling.In view of the obvious curvature difference between buildings and vegetation in the actual collected point cloud data,this paper proposes an improved fast segmentation method of 3D point cloud based on multi-layer lidar.After the initial segmentation of building facade is realized based on curvature segmentation,the weighted Euclidean clustering is used for the second iterative segmentation of the initial segmented point cloud,which can reduce the iterative process and avoid falling into local optimum.Through the unmanned platform data acquisition and processing experiments and public data experiments,the effectiveness of this method in the segmentation of building and vegetation is verified.According to the final segmentation results of the scene,the over segmentation rate,under segmentation rate and correct segmentation rate of the scene are counted,and compared with the traditional region growing segmentation algorithm.The results show that the algorithm has strong applicability and segmentation accuracy in different scenes.

Dynamic Programming Track-Before-Detect Algorithm Based on Local Gradient and Intensity Map

CHEN Ying, HAO Ying-guang, WANG Hong-yu, WANG Kun

Computer Science. 2022, 49 (8): 150-156. doi:10.11896/jsjkx.210700135

Abstract

PDF(3391KB) ( 665 )

References | Related Articles | Metrics

Aiming at the low detection probability of traditional DP-TBD algorithm in the infrared weak and small targets images with high background complexity and low SNR,a dynamic programming track-before-detect algorithm based on local gradient and intensity map is proposed.Firstly,the algorithm uses the local gradient and intensity algorithm (LIG) to preprocess the frame sequence images to obtain a new measurement model.Then,a new value function is constructed according to the correlation of the value function of adjacent frames.Finally,the DP-TBD is used to accumulate the new value function in multiple frames,so as to realize the track-before-detect of small and weak targets.Monte Carlo simulation experiment results show that when the signal-to-noise ratio is lower than 4dB,the detection probability of this algorithm is about 10% higher than that of the traditional DP-TBD algorithm and DBT algorithm.At the same time,in the real infrared weak and small target sequence image with complex background,the algorithm can also effectively track the target before detection under the condition of a constant false alarm rate,which improves the detection probability of the target.

Anchor Free Object Detection Algorithm Based on Soft Label and Sample Weight Optimization

WANG Can, LIU Yong-jian, XIE Qing, MA Yan-chun

Computer Science. 2022, 49 (8): 157-164. doi:10.11896/jsjkx.210600240

Abstract

PDF(3142KB) ( 605 )

References | Related Articles | Metrics

Similar to the Anchor Based object detection algorithm,the Anchor Free object detection algorithm based on feature points also encounters the problem of ambiguous samples when dividing positive and negative samples.That is,the training samples are divided either positive or negative according to the specific threshold and the position of feature points,and when the model trains samples whose feature point is near the borderline,it will incur great loss,which will make the model pay too much attention to these ambiguous samples and reduce the performance of the model.In view of this situation,this paper proposes to improve the Anchor Free object detection algorithm based on feature points from the three aspects of soft label,loss function and weight optimization.By making full use of Center Ness,the impact of ambiguous samples on network performance is mitigated and the accuracy of object detection is improved.To prove the effectiveness of the proposed method,the FCOS object detector is employed in the comparative experiments on the classical Pascal VOC and MS COCO datasets,respectively.Finally,the mAP of the detector on Pascal VOC dataset increases to 82.16%(an increase of 1.31%) and the AP_50-95 on MS COCO dataset increases to 35.8% (an increase of 1.3%).

Moderate Clothes-Changing Person Re-identification Based on Bionics of Binocular Summation

CHEN Kun-feng, PAN Zhi-song, WANG Jia-bao, SHI Lei, ZHANG Jin

Computer Science. 2022, 49 (8): 165-171. doi:10.11896/jsjkx.210600140

Abstract

PDF(2233KB) ( 478 )

References | Related Articles | Metrics

Moderate clothes-changing person re-identification is to find the same person from different camera scenes under the premise of considering the moderate change of clothes.The implementation of existing person re-identification methods is usually based on the assumption that the pedestrian’s clothing is invariant,so they rely on clothing-related features.Then,when the above assumptions are not valid,these methods are difficult to achieve the ideal recognition performance.Considering the important characteristic that pedestrian’s shape hardly change when the change of clothes is moderate,the moderate clothes-changing person re-identification is studied.Inspired by the binocular summation in biological vision system,a self-attention siamese network is proposed in this paper.Analogous to biological binocular information acquisition process,the network takes two types of images of the same pedestrian with different clothes as two branch inputs,and then achieves summation effect with siamese architecture.Subsequently,the contrastive learning and fusion learning of multiple features are carried out to obtain the pedestrian feature representation with identity discrimination.Finally,empirical studies show that the proposed method achieves best performance at present on clothes-changing person re-identification benchmark.

Non-local Attention Based Generative Adversarial Network for Video Abnormal Event Detection

SUN Qi, JI Gen-lin, ZHANG Jie

Computer Science. 2022, 49 (8): 172-177. doi:10.11896/jsjkx.210600061

Abstract

PDF(2305KB) ( 574 )

References | Related Articles | Metrics

As the uncertainty of abnormal events,the method of future frame prediction is chosen to detect abnormal events in video.The prediction model is trained with normal samples,so that the model can accurately predict the future frames without abnormal events.However,it cannot predict video frames with unknown events.Combining with apparent constraints and motion constraints,generative adversarial network is used to train the generator model for prediction.In order to reduce the loss of relative target features,a nonlocal attention Unet generator (NA-UnetG) model is proposed to improve the prediction accuracy of generator and the accuracy of abnormal video event detection.Experiments on datasets CUHK Avenue and UCSD Ped2 validate the effectiveness of the proposed method.The results show that the AUC of the proposed method is better than that of other methods,reaches 83.4% and 96.3%,respectively.

Study on Acceleration Algorithm for Raw Data Simulation of High Resolution Squint Spotlight SAR

GUO Zheng-wei, FU Ze-wen, LI Ning, BAI Lan

Computer Science. 2022, 49 (8): 178-183. doi:10.11896/jsjkx.210600066

Abstract

PDF(3290KB) ( 398 )

References | Related Articles | Metrics

Raw data simulation is the front-end work of synthetic aperture radar(SAR) system development,which is of great significance.For high resolution squint spotlight SAR,time-domain raw data simulation is usually used,but its simulation efficiency is very low.In order to realize the raw data simulation of high resolution squint spotlight SAR efficiently,an effective acceleration algorithm is proposed.To reduce the redundant computation and save memory,this algorithm combines the time-domain raw data simulation model and its signal characteristics to compensate the range cell migration(RCM) in the raw data simulation process of squint spotlight SAR.An adaptive data partitioning algorithm is adopted to compute the partitioned data in graphic processing unit(GPU), and the powerful computing capabilities of GPU is used to improve efficiency.Then the sub data blocks are transmitted and spliced in memory.The proposed algorithm improves the computational efficiency of time-domain raw data simulation,and solves the problems of huge volume of raw data,limited GPU memory and data transmission between video memory and memory.Experimental results of point targets and distributed targets show that the speedup ratio of this algorithm reaches 219.8,which verifies the effectiveness of the proposed method.

Multi-detector Fusion-based Depth Correlation Filtering Video Multi-target Tracking Algorithm

SHEN Xiang-pei, DING Yan-rui

Computer Science. 2022, 49 (8): 184-190. doi:10.11896/jsjkx.210600004

Abstract

PDF(3630KB) ( 440 )

References | Related Articles | Metrics

In the detection and tracking task,the detector has mis-detected and missed targets.For video multi-target tracking algorithms that rely on detection information,there will be a large number of false tracking targets and missed targets.Such missed and false targets will last for dozens of frames,resulting in reduced tracking accuracy.Due to this reason,a multi-detector fusion deep correlation filter video multi-target tracking algorithm is proposed.It uses the information of multiple detectors and proposes a new fusion mechanism to reduce the number of missed detections and false detections caused by a single detector,and break the performance limitations of a single detector,which makes the acquisition of new targets more reliable.On the other hand,the deep correlation filter algorithm ECO is used to track the targets one by one,and a series of improvements are proposed on the basis of the original algorithm ECO,which is more suitable for the video multi-target tracking task,and reduces the number of missed targets and identity tag jumps.Finally,experiments are carried out on the MOT17 data set,compared with the traditional video multi-target tracking method IOU17,MOTA of the proposed algorithm improves from 47.6 to 50.3.It is proved that this method has made great improvement in the research of multi-target tracking.

Methods in Adversarial Intelligent Game:A Holistic Comparative Analysis from Perspective of Game Theory and Reinforcement Learning

YUAN Wei-lin, LUO Jun-ren, LU Li-na, CHEN Jia-xing, ZHANG Wan-peng, CHEN Jing

Computer Science. 2022, 49 (8): 191-204. doi:10.11896/jsjkx.220200174

Abstract

PDF(4699KB) ( 6661 )

References | Related Articles | Metrics

Adversarial intelligent game is an advanced research in decision-making problem of intelligence cognitive.With the support of large computing power,game theory and reinforcement learning represented by counterfactual regret minimization and fictitious self-play respectively,are state-of-the-art approaches in searching strategies.However,the relationship between these two paradigms is not entirely explored.For adversarial intelligent game problems,this paper defines the connotation and extension of adversarial intelligent game,studies the development history of adversarial intelligent game,and summarizes the key challenges.From the perspectives of game theory and reinforcement learning,the models and algorithms of intelligent game are introduced.This paper conducts a comparative study from game theory and reinforcement learning,including the methods and framework,the main purpose is to promote the advance of intelligent game,and lay a foundation for the development of general artificial intelligence.

Review of Text Classification Methods Based on Graph Convolutional Network

TAN Ying-ying, WANG Jun-li, ZHANG Chao-bo

Computer Science. 2022, 49 (8): 205-216. doi:10.11896/jsjkx.210800064

Abstract

PDF(4281KB) ( 3785 )

References | Related Articles | Metrics

Text classification is a common task in natural language processing,in which there are a lot of research and progress based on machine learning and deep learning.However,these traditional methods can only process Euclidean spatial data,and cannot express the semantic information of the document effectively.To break the traditional learning mode,many recent studies start to use graphs to represent complicated relationships among entities in the document,and explore graph convolutional neural network for text representation.This paper reviews the text classification methods based on graph convolutional network.Firstly,the background and principle of graph convolutional network are summarized.Then,text classification methods based on graph convolutional network are described in detail according to different types of graph-based networks.Meanwhile,it analyzes the limi-tation of graph convolutional network in the depth of networks,and introduces the latest developments of deep networks in text classification.Finally,the classification performance of models involved in this paper is compared through some experiments,and the challenges and future research direction in this field are discussed.

Survey on Spiking Neural P Systems with Rules on Synapses

ZHANG Lu-ping, XU Fei

Computer Science. 2022, 49 (8): 217-224. doi:10.11896/jsjkx.220300078

Abstract

PDF(1737KB) ( 2507 )

References | Related Articles | Metrics

Membrane systems are a class of bio-inspired computing models,inspired by the structure and function of cells,tissue,organ and bio-systems.Spiking neural P systems with rules on synapses(SNPRS) are a type of membrane systems,inspired by the way that neurons communicate information.In SNPRS,each neuron is a basic unit for storing information,and each synapse is a medium for integrating and transmitting information.The whole system processes information in the distributed and parallel way.In this paper,we review the definition and related notions of SNPRS.Then,we introduce a few variants of SNPRS,and give a comparison among the variants of SNPRS.Furthermore,we provide results on the computation power of SNPRS(and their variants) working in different modes and on the application of the systems,such as solving NP-hard problems,implementing arithmetic operations,and breaking RSA.Additionally,some open problems are provided to suggest directions for further theore-tical as well as applicable research on SNPRS.

Design and Implementation of RPA System Based on UIA Interface

WANG Yan-song, QIN Yun-chuan, CAI Yu-hui, LI Ken-li

Computer Science. 2022, 49 (8): 225-229. doi:10.11896/jsjkx.211100046

Abstract

PDF(2054KB) ( 2781 )

References | Related Articles | Metrics

Robotic process automation(RPA)is one of the current research hotspots.RPA mainly uses Win32 interface to automate the operation of windows.This method requires to encapsulate the API,which is expensive to develop and can only identify UI elements designed based on Win32 API.At the same time,automation based on Win32 interface needs to simulate keyboard and mouse operations.Because these operations are based on a broadcast message mechanism,the response time is long.This paper proposes a solution for building RPA process automation application based on Microsoft's UIA technology.The solution uses UIA methods to automate UI elements.It can adapt to a wide range of UI program frameworks,including Win32,WPF,QT,Silverlight,etc.,and the development cost is low.At the same time,this method binds the message to the UI element,avoids the inefficient broadcast message mechanism,and improves the execution efficiency of RPA.Experimental results show that the execution time can be shortened by 55.67% on average compared with the keyboard and mouse method.

Text Classification Method Based on Information Fusion of Dual-graph Neural Network

YAN Jia-dan, JIA Cai-yan

Computer Science. 2022, 49 (8): 230-236. doi:10.11896/jsjkx.210600042

Abstract

PDF(2547KB) ( 2857 )

References | Related Articles | Metrics

Graph neural networks are recently applied in text classification tasks.Compared with graph convolution network,the text level graph neural network model based on message passing(MP-GNN) has the advantages of low memory usage and supporting online testing.However,MP-GNN model only builds a lexical graph using the word co-occurrence information, and the obtained information lacks diversity.To address this problem,a text classification method based on information fusion of dual-graph neural network is proposed.Besides preserving the original lexical graph built in MP-GNN,this method constructes the semantic graph based on the cosine similarity between pairs of words,and controls the sparsity of the graph through a threshold,which makes more effective use of the multi-directional semantic information of the text.In addition,the ability of direct fusion and attention mechanism fusion are tested to fuse the text representation learned on lexical graph and semantic graph.Experimental results on 12 datasets(R8,R52 and other datasets commonly used for text classification) show that the proposed model demonstrates an obvious improvement on performance of text classification compared with the SOTA(state-of-the-art) methods TextLevelGNN,TextING and MPAD.

Archimedes Optimization Algorithm Based on Adaptive Feedback Adjustment Factor

CHEN Jun, HE Qing, LI Shou-yu

Computer Science. 2022, 49 (8): 237-246. doi:10.11896/jsjkx.210700150

Abstract

PDF(2760KB) ( 2556 )

References | Related Articles | Metrics

Aiming at the problem of slow convergence speed of basic Archimedes optimization algorithm and is easy to fall into local optimum,this paper proposes an Archimedes optimization algorithm based on adaptive feedback adjustment factor.Firstly,initializing the population through the good point set to enhance the ergodicity of the initial population and improve the quality of the initial solution.Secondly,an adaptive feedback adjustment factor is proposed to balance the global exploration and local deve-lopment capabilities of the algorithm.Finally,the Levy rotation transformation strategy is proposed,to increase the diversity of the population and prevent the algorithm from falling into a local optimum.Comparative experiments of the proposed algorithm and mainstream algorithms are carried on 14 benchmark functions and some CEC2014 functions for 30 times.The optimization results of the algorithm on the function show that the average optimization accuracy,standard deviation and convergence curve of the proposed algorithm are better than that of the comparison algorithm.At the same time,Wilcoxon rank sum test is performed on 14 benchmark functions between the proposed algorithm and comparison algorithms.The test results show that the proposed algorithm is significantly different from comparison algorithms.It will be applied to the design of welded beams,which is 2% higher than the original algorithm,which verifies the effectiveness of the proposed algorithm.

Adaptive Reward Method for End-to-End Cooperation Based on Multi-agent Reinforcement Learning

SHI Dian-xi, ZHAO Chen-ran, ZHANG Yao-wen, YANG Shao-wu, ZHANG Yong-jun

Computer Science. 2022, 49 (8): 247-256. doi:10.11896/jsjkx.210700100

Abstract

PDF(4520KB) ( 3010 )

References | Related Articles | Metrics

At present,most multi-agent reinforcement learning(MARL) algorithms using the architecture of centralized training and decentralized execution(CTDE) have good results in homogeneous multi-agent systems.However,for heterogeneous multi-agent systems composed of different roles,there is always the problem of credit assignment,which makes it difficult for agents to learn effective cooperation strategies.To tackle the above problems,an adaptive reward method with end-to-end cooperation based on multi-agent reinforcement learning is proposed.It can promote the cooperation between agents.First,a batch regularization network is proposed.It uses a graph neural network to model the cooperative relationship of heterogeneous multi-agents.And it uses the attention mechanism to calculate the weight of key information.Also,it uses the batch regularization method to generate feature vectors.Besides,it guides the algorithm to learn in the right direction,thereby effectively improving the performance of heterogeneous multi-agent cooperative strategy generation.Second,an adaptive intrinsic reward network based on the actor-critic method is proposed.It can convert sparse rewards into dense rewards,which can guide agents to generate cooperative strategies according to the situation on the field.Through experiments,compared with the current mainstream multi-agent reinforcement learning algorithms,the proposed method has achieved significantly good results in the “cooperative-game” scenario.In addition,the visual analysis of the strategy-reward-behavior correlation further verifies the effectiveness of the proposed method.

Construction and Multi-feature Fusion Classification Research Based on Multi-scale Sparse Brain Functional Hyper-network

LI Yao, LI Tao, LI Qi-fan, LIANG Jia-rui, Ibegbu Nnamdi JULIAN, CHEN Jun-jie, GUO Hao

Computer Science. 2022, 49 (8): 257-266. doi:10.11896/jsjkx.210600094

Abstract

PDF(3889KB) ( 2700 )

References | Related Articles | Metrics

Brain functional hypernetworks have been successfully utilized in the diagnosis of brain diseases.In the previous study,the different hyper-edge generation method was mainly used to improve the construction of the hyper-network,which ignored the influence of different nodes definitions on the brain functional hyper-network topology.Therefore,in light of this problem,it is proposed to construct a brain functional hyper-network based on parcellation of different scales,so as to analyze its impact on brain functional hyper-network topology and classification performance.Specifically,firstly,based on the anatomical automatic labeling atlas,the brain was segmented by the method of clustering algorithm and the random dynamic seed point;secondly,based on the average time series obtained under each node scale,the brain functional hyper-network was constructed by the LASSO method respectively;then multiple sets of local features (node degree,shortest path,clustering coefficient) were extracted,and non-parametric tests and correlation-based methods were used to select features with difference;finally,support vector machine was adopted to build classification model.The classification results show that as the size of nodes increases,the classification accuracy of the constructed brain functional hyper-network is higher.When the node scale is 1501,the classification accuracy can reach 95.45%.Meanwhile,the classification accuracy of multi-scale fusion is better than that of any scale,which indicate different node definitions will affect the topology of the brain functional hyper-network.In future research,besides focusing to the construction method of the hyper-edge,the choice of brain parcellation scheme needs more attention in hyper-network.Moreover,combining multi-scale features can supplement more classification information to enhance the classification performance of depression and normal control.In addition,regardless of the size of the node,the classification performance of multiple sets of local properties is better than that of a single type of properties,which illustrates multiple sets of local property can make up for the missing information of a type of single feature,thereby discovering more brain disease biological markers.While effectively representing the brain functional hyper-network,it is also necessary to quantify the brain functional hyper-network topology information from multiple angles,so that the ability to characterize differences between groups can be enhanced,and the ability to diagnose and predict brain diseases can be improved.

KPCA Based Novelty Detection Method Using Maximum Correntropy Criterion

LI Qi-ye, XING Hong-jie

Computer Science. 2022, 49 (8): 267-272. doi:10.11896/jsjkx.210700175

Abstract

PDF(2031KB) ( 2654 )

References | Related Articles | Metrics

Novelty detection is an important research issue in the field of machine learning.Till now,there exist lots of novelty detection approaches.As a commonly used kernel method,kernel principal component analysis(KPCA)has been successfully applied to deal with the problem of novelty detection.However,the traditional KPCA based novelty detection method is very sensitive to noise.If there exist noise in the given training samples,the detection performance of KPCA based novelty detection method may be decreased.To enhance the anti-noise ability of KPCA based novelty detection method,a maximum correntropy criterion(MCC)based novelty detection method is proposed.Correntropy in information theoretic learning is utilized to substitute the ₂-norm based measure in KPCA based novelty detection method.By adjusting the width parameter of the correntropy function,the adverse effect of noise can be alleviated.The half-quadratic optimization technique is used to solve the optimization problem of the proposed method.The local optimal solution can thus be obtained after a few iterations.Moreover,the algorithmic description of the proposed method is provided,and the computational complexity of the corresponding algorithm is analyzed.Experimental results on the 16 UCI benchmark data sets demonstrate that the proposed method obtains better anti-noise and generalization performance in comparison with the other four related approaches.

Deformable Graph Convolutional Networks Based Point Cloud Representation Learning

LI Zong-min, ZHANG Yu-peng, LIU Yu-jie, LI Hua

Computer Science. 2022, 49 (8): 273-278. doi:10.11896/jsjkx.210900023

Abstract

PDF(2340KB) ( 2531 )

References | Related Articles | Metrics

Although the sparseness and irregularity of point cloud data have been successfully solved by deep neural networks.However,how to learn the local features of point clouds is still a challenging problem.Existing networks for point cloud representation learning have the problem of extracting features independently between points and points.To this end,a new spatial graph convolution is proposed.Firstly,an adaptive hole K-nearest neighbor algorithm is proposed when constructing the graph structure to maximize local topo-logical structure information.Secondly,the angle feature between each edge of the convolution kernel and the receptive field map is added to the convolution,which ensures more discriminative feature extraction.Finally,in order to make full use of local features,a novel graph pyramid pooling is proposed.This algorithm is tested on the standard public data sets ModelNet40 and ShapeNet,and the accuracy is 93.2% and 86.5% respectively.Experimental results show that the proposed algorithm is at a leading level in point cloud representation learning.

Survey of Social Network Public Opinion Information Extraction Based on Deep Learning

WANG Jian, PENG Yu-qi, ZHAO Yu-fei, YANG Jian

Computer Science. 2022, 49 (8): 279-293. doi:10.11896/jsjkx.220300099

Abstract

PDF(1927KB) ( 1870 )

References | Related Articles | Metrics

With the rapid development of social media platforms,public opinion information can be widely disseminated in a very short period of time.If the information of public opinion is not managed and controlled,it will pose a great threat to the network environment and even the social environment.Information extraction technology has become the first and the most significant step in public opinion analysis and management due to its semantization and accuracy.Over the last few years,with the development of deep learning,its ability to automatically learn potential features and combine these features has dramatically improved the accuracy of each sub-task of information extraction.This paper systematically composes and summarizes the methods of extracting information by combining the characteristics of social media public opinion and deep learning technology.Firstly,we sort out the organization of public opinion information in social networks,elaborate the framework and evaluation indexes of public opinion information extraction.Then we conduct a comprehensive review and analysis of existing deep learning-based public opinion information extraction models,discuss the applicability and limitations of existing methods.Finally,the future research trends is prospected.

Survey of Ethereum Smart Contract Fuzzing Technology Research

HUANG Song, DU Jin-hu, WANG Xing-ya, SUN Jin-lei

Computer Science. 2022, 49 (8): 294-305. doi:10.11896/jsjkx.220500069

Abstract

PDF(1844KB) ( 1518 )

References | Related Articles | Metrics

Smart contracts running on the blockchain platform completethe establishment and automatic execution of a greements between different participants,and also manage a large number of digital assets.The frequent exposure of smart contract loopholes has caused incalculable economic losses.Fuzzing is an effective dynamic vulnerability detection technique that has been applied to smart contract security research.This paper analyzes the problem of insufficient summarization of smart contract fuzzing in existing review work,and proposes a basic framework for smart contract fuzzing.Taking Ethereum smart contracts as an example,which are currently the most widely studied in smart contract security,the account mechanism and transaction structure closely related to smart contracts are introduced,and the characteristics of smart contracts that are different from traditional programs are summarized.The vulnerabilities of smart contracts are expounded,and the vulnerabilities covered by these smart contract fuzzing techniques are compared.Furthermore,the input generation of the existing smart contract fuzzing technology is analyzed from the aspects of single transaction and transaction sequence.The input mutation is summarized from the functional level,transaction level and transaction sequence level.The use of test oracles for existing smart contract fuzzing techniques is briefly described.In addition,the corresponding technical evaluation indicators are also summarized.Finally,the problems faced by smart contract fuzzing are proposed,and the future research directions are prospected.

Authentication and Key Agreement Protocol for UAV Communication

JIAN Qi-rui, CHEN Ze-mao, WU Xiao-kang

Computer Science. 2022, 49 (8): 306-313. doi:10.11896/jsjkx.220200098

Abstract

PDF(2107KB) ( 821 )

References | Related Articles | Metrics

In order to achieve the requirement for security and lightweight in unmanned aerial vehicle(UAV)communication,two authentication and key agreement protocols targeted for UAVs with different computational performance are proposed,including an ECC based protocol,DroneSec,and a symmetric cipher based protocol,DroneSec-lite.The two protocols achieve secure mutual authentication and key configuration between ground stations and UAVs.DroneSec protocol achieves relatively low computational overhead while ensuring forward secrecy through combining ECDH and MAC,which is suitable for relatively high-performance platforms.DroneSec-lite protocol achieves extremely low computational overhead through using only symmetric ciphers,which is suitable for low-performance platforms.The security of the proposed protocols under the enhanced Dolve-Yao model is verified using ProVerif,a formal protocol verification tool,and the performance of the protocols is analyzed in the simulation environment.The results show that it is superior to existing protocols in terms of computation overhead,communication overhead and security.

Network Traffic Anomaly Detection Method Based on Multi-scale Memory Residual Network

WANG Xin-tong, WANG Xuan, SUN Zhi-xin

Computer Science. 2022, 49 (8): 314-322. doi:10.11896/jsjkx.220200011

Abstract

PDF(3013KB) ( 513 )

References | Related Articles | Metrics

Network traffic anomaly detection based on deep learning usually has the problems of poor adaptability to real-world environments,limited representation ability and week generalization ability.From the perspective of these problems,a network traffic anomaly detection method based on multi-scale memory residual network is proposed.Based on the analysis of high-dimensional feature space distribution,this paper demon-strates the validity of the approach to network traffic data preprocessing.Combining multi-scale one-dimensional convolution and long short-term memory network,the representation ability is enhanced by deep learning classifiers.To make the network traffic anomaly detection accurate and efficient,by the idea of residual network,the deep feature extraction is implemented,the problems of vanishing/exploding gradients,the over-fitting and network degradation are prevented,and the convergence speed of the model is accelerated.The visualizations of data preprocessing result suggest that,compared with standardization,normalization has better capability to separate the abnormal traffic data from the normal traffic data.The result of validity verification and performance evaluation experiment reveal that,by inserting identity mapping,the convergence speed of the model can be accelerated,and the network degradation problem can be efficiently addressed.The result of contrast experiment indicates the one-dimensional convolution and long short-term memory network can reinforce the representation and generalization ability of our model,and the performance metrics of our model is better than that of the current deep learning model.

Class Discriminative Universal Adversarial Attack for Text Classification

HAO Zhi-rong, CHEN Long, HUANG Jia-cheng

Computer Science. 2022, 49 (8): 323-329. doi:10.11896/jsjkx.220200077

Abstract

PDF(2573KB) ( 446 )

References | Related Articles | Metrics

The definition of universal adversarial attack is that the text classifiers can be successfully fooled by a fixed sequence of perturbations appended to any inputs.But textual examples from all classes are indiscriminately attacked by the existing UAA,which is easy to attract the attention of the defense system.For more stealth attack,a simple and efficient class discriminative universal adversarial attack method is proposed,which has an obvious attack effect on textual examples from the targeted classes and limited influence on the non-targeted classes.In the case of white-box attack,multiple candidate perturbation sequences are searched by using the average gradient of the perturbation sequence in each batch.The perturbation sequence with the smallest loss is selected for the next iteration until no new perturbation sequence is generated.Comprehensive experiments are conducted on four public Chinese and English datasets and TextCNN,BiLSTM to evaluate the effectiveness of the proposed method.Experimental results show that the proposed attack method can discriminatively attack the targeted and non-targeted classes,and has certain transferability.

Rumor Detection Model Based on Improved Position Embedding

JIANG Meng-han, LI Shao-mei, ZHENG Hong-hao, ZHANG Jian-peng

Computer Science. 2022, 49 (8): 330-335. doi:10.11896/jsjkx.210600046

Abstract

PDF(2291KB) ( 569 )

References | Related Articles | Metrics

With the rise of online social networks,the way people disseminate and obtain information has changed drastically.While social media facilitates people’s lives,it also accelerates the generation and spread of rumors.For this reason,detect rumors accurately and efficiently becomes an urgent problem to be solved.In order to improve the accuracy of rumor detection,the rumor detection model based on the global-local attention network has been improved.Taking into account the influence of the positional relationship between words in the text on rumor detection,a new relative position encoding method is introduced to improve the local feature extraction module of the original model.This method can more accurately extract and aggregate the semantic information and location information of the text in the rumor,and obtain better text features that distinguish between rumors and non-rumors.The combination of features and global features describing forwarding behavior improves the detection effect of rumors.Experimental results show that,compared with other mainstream detection methods,the F1 value of the proposed method can reach 95.0% on the Microblog data set,which has a better detection effect.

Study on Malware Classification Based on N-Gram Static Analysis Technology

ZHANG Guang-hua, GAO Tian-jiao, CHEN Zhen-guo, YU Nai-wen

Computer Science. 2022, 49 (8): 336-343. doi:10.11896/jsjkx.210900203

Abstract

PDF(2915KB) ( 673 )

References | Related Articles | Metrics

In order to solve the problem of low accuracy of malware classification,this paper proposes a research on malware classification based on N-Gram static analysis technology.Firstly,the N-Gram method is used to extract the byte sequence of length 2 from the malware samples.Secondly,according to the extracted features,KNN,logistic regression,random forest and XGBoost are used to train the malware classification model based on machine learning.Thirdly,the confusion matrix and logarithmic loss function are used to evaluate the malware classification model.Finally,the malware classification model is trained and tested in the Kaggle malware data set.Experimental results show that the accuracy rates of the malware classification models of XGBoost and random forest reach 98.43% and 97.93%,and the Log Loss values are 0.022240 and 0.026946,respectively.Compared with the existing methods,the proposed method can classify malware more accurately and protect computer system from malware attack.

Compressed Image Encryption Scheme Based on Dual Two Dimensional Chaotic Map

ZHOU Lian-bing, ZHOU Xiang-zhen, CUI Xue-rong

Computer Science. 2022, 49 (8): 344-349. doi:10.11896/jsjkx.210700235

Abstract

PDF(2816KB) ( 423 )

References | Related Articles | Metrics

In order to improve the security and encryption speed of chaotic encryption system,a lightweight compression image encryption scheme based on chaotic mapping is proposed.Firstly,in order to speed up the encryption process and destroy the correlation between pixels,the digital image is divided into image blocks and compressed by processing in the frequency domain.Then,2D Logistic chaotic mapping is implemented to rearrange and transpose image pixels in the key generation,scrambling and replacement stages.In addition,to further improve the security level,in the diffusion stage,another chaotic map 2D Henon map is used to change the pixel value of the scrambled image.Performance evaluation results show that the proposed scheme can meet the security requirements of image encryption,and the key space is large.Compared with other excellent schemes,this scheme can better resist statistical and differential attacks,and it is highly sensitive to the change of the generated key space.Its key space size is 256bits,the NPCR value is about 99%,and UACI value is more than 15%.