Computer Science

Key Value Storage Technology Based on LSM-tree:A Survey

LYU Meng, HUA Wendi, XIE Ping

Computer Science. 2023, 50 (8): 1-15. doi:10.11896/jsjkx.220900178

Abstract

PDF(1995KB) ( 1404 )

References | Related Articles | Metrics

Key-value storage is the simplest form of database organization and it plays a key role in data-intensive application scenarios.With the increasing demand for timely data analysis,good system performance becomes more and more important.At present,the storage engine of most key-value storage systems is the log-structured merge tree(LSM-tree).Because of its excellent write performance,the LSM-tree is widely used in the write intensive scenes and the storage layer of modern NoSQL system.Compared to the traditional B-tree,LSM-tree adopts sequential write access mode.At the same time,it uses memory buffer to batch new write threads,so it has greater write advantages.Nevertheless,repeated reading and writing of data and unnecessary compression operations lead to the problem of read and write amplification of the LSM-tree.Finally,these problems seriously affect the performance of the system,especially in the data-intensive application scenarios.Nowadays,researches have make great efforts to solve the problems.Firstly,this paper investigates various factors that affect the performance of the LSM-tree,collects a lot of literature on improving the performance of LSM tree-based key-value systems,organizes and categorizes them.Then it discusses their advantages and tradeoffs to enable readers to understand LSM tree-based storage technologies and their optimization strategies.Finally,several representative LSM tree-based key-value storage technologies are surveyed and some potential future research directions are discussed.

Maximum Influential Community Search in Heterogeneous Information Network

DU Ming, YANG Wen, ZHOU Junfeng

Computer Science. 2023, 50 (8): 16-26. doi:10.11896/jsjkx.220600262

Abstract

PDF(3442KB) ( 1437 )

References | Related Articles | Metrics

Heterogeneous information network can effectively model data systems,which have diverse object types and complex interactions.Research on community search based on heterogeneous information networks usually builds community models centered on vertex type,minimum degree and network structure,then the cohesive subgraph is queried.However,there are two pro-blems in the existing researches:1)the influence value,another natural attribute hidden in networks is not considered;2)the user'srequirement for the upper limit of the query result scale is ignored too,resulting in the query result do not match user's expectation.Therefore,this paper studies the heterogeneous information networks combined with influence value,and proposes a combined constraint model as a measure of community cohesion for such networks.To solve the community search problems based on the combined constraint model,this paper proposes two search algorithms optimized by preprocessing and pruning strategies.Finally,the effectiveness and efficiency of our method are verified on 8 real data sets.

Quantum Prototype Clustering

LIU Xiang, ZHU Jing, ZHONG Guoqiang, GU Yongjian, CUI Liyuan

Computer Science. 2023, 50 (8): 27-36. doi:10.11896/jsjkx.220600124

Abstract

PDF(3993KB) ( 1252 )

References | Related Articles | Metrics

Quantitative reconstruction of classical machine learning algorithms is one of the significant research directions in the field of quantum machine learning.The quantitative implementation of clustering algorithm,which has been widely used in the machine learning area,is worth studying.Most of the current quantum machine learning algorithms suffer from the difficulty of reproduction and the difficulty of forming direct comparisons with classical algorithms.To address these problems,this paper proposes the quantum prototype clustering algorithm that can be easily deployed on existing general-purpose quantum computing devices.Combining the rotation property of single quantum bit(qubit),and finding the feature mapping method with minimal information loss,the single qubit rotation is created using two-dimensional feature data.Then,based on the properties of multi-qubit entanglement and the collapse of the entangled system,a quantum circuit is designed for generating a specific quantum entangled system and measuring the collapsed result of entangled system.Based on the relationship between the rotation angle of controlled qubits in the entangled system and the collapse result of the entangled system,and combined with the definition of Minkowski distance,a quantum distance for evaluating the similarity of input samples is then derived.Both the quantum distance calculation module and its counterpart in classic computer have the same forms of input and output,so that the latter can be replaced by the former without modification,hence the prototype clustering algorithm is quantitatively reconstructed into QPC.Several replicated experiments on multiple publicly available datasets from kaggle and scikit-learn show that the QPC performs similarly to classic prototype clustering algorithm in terms of many evaluation metrics,such as the mean sample-centroid distance.

Study on Multimodal Online Reviews Helpfulness Prediction Based on Attention Mechanism

ZHANG Yian, YANG Ying, REN Gang, WANG Gang

Computer Science. 2023, 50 (8): 37-44. doi:10.11896/jsjkx.220600204

Abstract

PDF(3830KB) ( 1455 )

References | Related Articles | Metrics

In the e-commerce era,online reviews are regarded as important product evaluations,which profoundly influence consumers' decision-making process.However,the exponentially increasing number of reviews and unstructured review data pose challenges to feature selection and accuracy improvement of review helpfulness prediction.In addition,current research mainly focuses on shallow features and feature extraction of review texts,the image information contained in review photos is often ignored.Besides,multi-modal information such as review text,photos,and shallow features needs to be refined and fused by app-lying multi-modal fusion methods.Based on these,this paper regards review photos and review text as a latent feature affecting the helpfulness of online reviews,and designs a shallow feature set according to the KAM knowledge adoption theory.For the data of three modalities,a deep prediction model,i.e.,three-modal review helpfulness prediction based on co-attention mechanism(TMCAM) is proposed,which can achieve the interaction and fusion of cross-modal information.The superior performance of the TMCAM model is tested through experiments,and it is proved that the complementation of image and text information can achieve better results than single modal information.Besides,shallow features can help predict the reviews helpfulness.Moreover,compared with simple modal features splicing,using collaborative attention mechanism for cross-modal information interaction helps to improve the perception of reviews helpfulness.

Traffic Data Restoration Method Based on Tensor Weighting and Truncated Nuclear Norm

WU Jiangnan, ZHANG Hongmei, ZHAO Yongmei, ZENG Hang, HU Gang

Computer Science. 2023, 50 (8): 45-51. doi:10.11896/jsjkx.220600160

Abstract

PDF(3627KB) ( 1204 )

References | Related Articles | Metrics

The problem of missing data seriously affects a series of activities in intelligent transportation systems,such as monitoring traffic dynamics,predicting traffic flow,and deploying traffic planning through data.Therefore,a traffic flow data reconstruction model WLRTC-TTNN(low rank tensor completion of weighted and truncated nuclear norm)combined with weighted and truncated nuclear norm is proposed by using the low-rank tensor completion framework based on tensor singular value decomposition,which can effectively repair the missing spatio-temporal traffic data.The truncated nuclear norm of the tensor is used as a convex proxy for tensor rank minimization instead of tensor rank minimization,which preserves the main feature information inside the spatio-temporal traffic data,and further optimizes the model by penalizing smaller singular values according to the gene-ralized singular value threshold theory,and finally the WLRTC-TTNN algorithm is implemented using the alternating multiplier method.Experiments are conducted on two publicly available spatio-temporal traffic datasets selected with different missing scenarios and missing rates,and the results show that the complementary performance of WLRTC-TTNNN is better than that of other baseline models,and the overall complementary accuracy improves by 3%~37%,and the complementary effect is more stable in extreme missing scenarios.

Data Completion of Air Quality Index Based on Multi-dimensional Sparse Representation

CAI Qiquan, LU Juhong, YU Zhiyong, HUANG Fangwan

Computer Science. 2023, 50 (8): 52-57. doi:10.11896/jsjkx.220500277

Abstract

PDF(2183KB) ( 1142 )

References | Related Articles | Metrics

In recent years,air pollution has become increasingly serious and become one of the risk factors affecting people's health.The air quality index(AQI) can provide the government with the laws of atmospheric environment changes,and can also be used for air pollution control.However,the data is inevitably missing in the process of collection,which leads to the difficulty of data mining.However,given the poor performance of existing completion methods under a high miss rate,this paper transforms the missing-matrix-completion problem into a sparse-matrix-reconstruction problem and designs a data completion method based on multi-dimensional sparse representation.The method first uses the training data to simulate various random missing cases for over-complete dictionary learning.Then,the sparse representation of the matrix with missing values is obtained by using the upper part of the learned dictionary.Finally,the sparse representation is combined with the lower part of the dictionary to obtain the reconstructed estimation matrix.Experimental results show that the proposed algorithm is superior to the traditional matrix method in the completion of multi-dimensional time series of AQI,especially in the case of serious missing.

OJ Exercise Recommendation Model Based on Deep Reinforcement Learning and Program Analysis

JIN Tiancheng, DOU Liang, ZHANG Wei, XIAO Chunyun, LIU Feng, ZHOU Aimin

Computer Science. 2023, 50 (8): 58-67. doi:10.11896/jsjkx.220600260

Abstract

PDF(2300KB) ( 1229 )

References | Related Articles | Metrics

At present,there are a large number of exercises on the existing programming Online Judge systems(OJ),which makes it difficult for students to quickly find suitable exercises according to their own knowledge level and learning demand.Therefore,it is necessary to design a model to recommend suitable exercises to students.However,due to uniqueness of OJ and complexity of programming ability evaluation,existing recommendation model can not complete OJ exercise recommendation task well,the main problems include:OJ exercises' lack of knowledge label and unique proposition style make it difficult for existing models to mine correlation between exercises; actual correctness of the program submitted by student is inconsistent with OJ judgement result,which leads to deviation of students' knowledge state estimated by models; existing models are difficult to provide exercises that increase students' programming ability most significantly.Based on this,this paper proposes an OJ exercise recommendation model based on deep reinforcement learning and program analysis.Firstly,analyzing optimal solution of exercises to mine correlations between exercises.Then,comparing the similarity between programs submitted by students and optimal solution of exercises to check actual correctness of the programs submitted by students,so that knowledge state of students can be estimated more accurately.Finally,using deep reinforcement learning technology,taking knowledge tracking model as student simulator and treating student simulator's performance difference on all the exercises before and after answering exercises provided by exercise re-commendation model as reward,so that exercise recommendation model can learn which exercise is able to improve the students' programming ability to the greatest extent,and recommend such exercises to students.This paper conducts extensive experiments on two datasets CodeForces and Libre of the well-known OJ system,and experimental results show that the proposed model can achieve higher performance than the state-of-the-art recommendation models.

Review of Talking Face Generation

SONG Xinyang, YAN Zhiyuan, SUN Muyi, DAI Linlin, LI Qi, SUN Zhenan

Computer Science. 2023, 50 (8): 68-78. doi:10.11896/jsjkx.221000031

Abstract

PDF(3632KB) ( 523 )

References | Related Articles | Metrics

Talking face generation is a popular research direction in the field of visual generation,which aims to generate realistic speaker videos based on multimodal input data.Talking face generation has broad application prospects in video media,game animation and Internet-related industries,and it could also provide data support for the research of tasks such as lip reading recognition,fake identification and digital human generation.The existing mainstream methods have been able to achieve talking face generation with personalized attributes and audio-visual synchronization,but they fail to meet the requirements of emerging application scenarios such as virtual reality,man-machine interaction and metaverse.Sothestudyof talking face generation is of great significance for promoting the development of related industries.This paper sorts out and summarizes the research status of tal-king face generation.First,itelaborates the research background and related technologies of talking face generation,then introduces the mainstream generation methods in recent years according to the method classification,sorts out the audio-visual datasets and evaluations commonly used in the research,and finally summarizes the problems in the existing methods,and analyzes the potential research direction of talking face generation in the future.

Survey of Rotating Object Detection Research in Computer Vision

WANG Xu, WU Yanxia, ZHANG Xue, HONG Ruize, LI Guangsheng

Computer Science. 2023, 50 (8): 79-92. doi:10.11896/jsjkx.221000148

Abstract

PDF(3474KB) ( 839 )

References | Related Articles | Metrics

Traditional object detector locates objects by horizontal bounding box(HBB),which often have low accuracy and poor generalization ability when detecting objects with arbitrary orientation angle,dense distribution,large aspect ratio and complex background.The above problems can be effectively solved by adding rotating target boxes with different rotation angles in the bounding boxes.This method is widely used in the fields of remote sensing images,scene text images,shelf goods images and other target detection,and has important research value.Most of the current works aim at constructing different models for rotating object detection,and there are fewer review works for summarizing and analyzing existing models in depth.Therefore,this paper provides a detailed review of existing research results on rotating object detection.Firstly,according to the current popular way of target box characterization,the target boxes are classified into three types of oriented bounding box(OBB),quadrilateral bounding boxes(QBB) and point set for generalized analysis,and simultaneously compare the advantages and disadvantages,network structures and performance of different rotating object detection algorithms.Secondly,the commonly used rotating object detection datasets and performance evaluation metrics are analyzed.Finally,the problems in the current study are briefly summarized and discussed,and the future development trend is prospected.

Adaptive Object Counting Model for Aerial Imagery

WEI Chang, GUAN Jihong, ZHANG Yichao, LI Wengen

Computer Science. 2023, 50 (8): 93-98. doi:10.11896/jsjkx.220600258

Abstract

PDF(3132KB) ( 243 )

References | Related Articles | Metrics

Object counting aims to obtain the number of specific types of objects such as vehicles,buildings,people contained in a given image,which is of great significance to urban planning,emergency response,national security,etc.The current object coun-ting task mainly relies on the images taken by low-altitude cameras,and there are obvious problems such as the object being easily occluded and the small counting space range.Widespread use of high-definition aerial remote sensing imagery makes it possible to count objects in large areas.However,the object counting task for aerial images has challenges such as large differences in object scales,dense distribution,and uncertain orientation.Existing object detection counting models and regression counting models based on low-altitude images are not suitable for object counting in aerial images.To solve this problem,this paper proposes an adaptive object counting model for aerial images.Firstly,the geometric adaptive Gaussian convolution method is used to solve the problem of object scale variation.Then,the image loss judgment method based on structural similarity is used to solve the pro-blem of poor counting stability of object dense regions.Experimental analysis shows that the proposed model can achieve better object count accuracy than the benchmark model.

Image Captioning Optimization Strategy Based on Deep Learning

ZHOU Ziyi, XIONG Hailing

Computer Science. 2023, 50 (8): 99-110. doi:10.11896/jsjkx.230200091

Abstract

PDF(3037KB) ( 330 )

References | Related Articles | Metrics

Image captioning aims to describe image content with grammatically correct sentences and automatically generate text.Image captioning involves computer vision and natural language processing,which is a classic task in multimodal field.In recent years,a large number of studies have begun to focus on image captioning,a multimodal task that combines vision and language,and has achieved many breakthrough results.Most of the existing surveys on image captioning take technology as the core and analyze from the perspective of classification.Considering that image captioning based on deep learning has become the mainstream research method at present,and its essence is an image-to-sequence problem,this paper takes visual input subtasks and language output subtasks as the theme,takes optimization strategy as the core.The optimization logic and technical development trend of these two subtasks are compared and analyzed.The existing challenges and task variations of image captioning are discussed.Finally,the optimization strategy and development direction of image captioning based on deep learning are expected to be further clarified.

Non-autoregressive Transformer Chinese Speech Recognition Incorporating Pronunciation- Character Representation Conversion

TENG Sihang, WANG Lie, LI Ya

Computer Science. 2023, 50 (8): 111-117. doi:10.11896/jsjkx.220600144

Abstract

PDF(1726KB) ( 264 )

References | Related Articles | Metrics

The Transformer based on self-attention mechanism shows powerful model performance in speech recognition tasks,where the non-autoregressive Transformer automatic speech recognition model has a faster decoding speed compared with the autoregressive model.However,the increase in speech recognition speed causes a larger decrease in accuracy.To improve the accuracy of the non-autoregressive Transformer speech recognition model,the frame information merging based on connectionist temporal classification(CTC) is introduced firstly,which fuses the speech high-dimensional representation in the frame width range to improve the problem of incomplete feature information in the non-autoregressive Transformer decoder input sequences.Secon-dly,pronunciation-character representation conversion is performed on the model output,and the pronunciation representation is converted into an output containing more character features by fusing contextual information on the pronunciation features of the decoder output,thus improving the recognition error problem of the model with different characters in the same pronunciation.Experiments on the Chinese speech dataset AISHELL-1 show that the proposed model achieves a recognition speed of real time factor(RTF) 0.0028 and recognition accuracy of 8.3% character error rate(CER),demonstrating strong competitiveness among many mainstream Chinese speech recognition algorithms.

Vietnamese Speech Synthesis Based on Transfer Learning

YANG Lin, YANG Jian, CAI Haoran, LIU Cong

Computer Science. 2023, 50 (8): 118-124. doi:10.11896/jsjkx.220600045

Abstract

PDF(3116KB) ( 315 )

References | Related Articles | Metrics

Vietnamese is the official language of the Socialist Republic of Vietnam.It belongs to the Vietnamese branch of the Viet Muang language family of the South Asian language family.In recent years,deep learning-based speech synthesis has been able to synthesize high-quality speech.However,these methods often rely on large-scale high-quality speech training data.An effective way to solve the problem of insufficient data for some low-resource non-lingua franca speech training is to adopt a transfer learning method and borrow other high-resource lingua franca speech data.Under the condition of low resources,with the goal of improving the quality of Vietnamese speech synthesis,the end-to-end speech synthesis model Tacotorn2 is selected as the baseline model,and the effects of different source languages,different text character embedding methods and transfer learning methods on the effect of speech synthesis are studied by transfer learning methods.Then,from both subjective and objective aspects,the speech synthesized by the various models described in this paper is evaluated.Experimental results show that the transfer learning system based on English phonetic module embedding＋Vietnamese phonology embedding method has achieved good results in synthesizing naturally understandable Vietnamese speech,and the MOS score of synthetic speech can reach 4.11,which is much higher than the 2.53 of the baseline system.

Lightweight Multi-view Stereo Integrating Coarse Cost Volume and Bilateral Grid

ZHANG Xiao, DONG Hongbin

Computer Science. 2023, 50 (8): 125-132. doi:10.11896/jsjkx.220600046

Abstract

PDF(2803KB) ( 247 )

References | Related Articles | Metrics

In order to tackle the problems of large memory consumption,poor real-time performance and poor reconstruction quality for low-textured areas of multi-view stereo reconstruction algorithm basedon deep learning,this paper proposes a lightweight cascade MVS reconstruction network based on bilateral grid and fused cost volume.Firstly,it builds the cost volume upsampling module based on learned bilateral grid,which can efficiently restore the low-resolution cost volume to the high-resolution cost volume.Then the dynamic region convolution and coarse cost volume fusion module are used to improve the network's ability to extract the feature of the challenging area and to perceive the global and structural information of the scene.Experimental results show that our method achieves competitive results on DTU dataset and tanks and temples benchmark,and is significantly better than other methods in memory consumption and inference speed.

Remote Sensing Image Pan-sharpening Method Based on Generative Adversarial Network

YAN Yan, SUI Yi, SI Jianwei

Computer Science. 2023, 50 (8): 133-141. doi:10.11896/jsjkx.220600065

Abstract

PDF(3399KB) ( 256 )

References | Related Articles | Metrics

Remote sensing image pan-sharpening methods are generally based on Wald protocol,resulting in blurred texture details,colors and ambiguous boundaries of the reconstructed images.To solve the problem,a remote sensing image pan-sharpening method based on generative adversarial networks(GAN),PAN-GAN,is proposed in this paper.The multispectral image is employed as the reference image.The grayscale reference image is applied to simulate the panchromatic image and the blurred reference image is adpoted as input of the generator.The generator extracts the texture details of the grayscale reference image and spectral features of the blurred reference image for the fusion reconstruction.Meanwhile,the perceptual loss is introduced to optimize the reconstruction results with adversarial loss and pixel loss,so that the reconstructed images have spectral and texture detail features closer to the reference image.Experiments are carried out on the datasets of three remote sensing satellites including QuickBird,GaoFen-2 and WorldView-2.The results show that the reconstructed images obtained by PAN-GAN have more realistic spectral and spatial texture details compared with common methods.The usage of grayscale reference images can significantly improve the performance of the original method,and the average grayscale improvement is the most obvious.The perceptual loss can further optimize the reconstruction results and verify the effectiveness of the proposed method.

Survey of Domain Adaptive Methods with Knowledge Integrating

CUI Fuwei, WU Xuanxuan, CHEN Yufeng, LIU Jian, XU Jin'an

Computer Science. 2023, 50 (8): 142-149. doi:10.11896/jsjkx.220800040

Abstract

PDF(1515KB) ( 377 )

References | Related Articles | Metrics

When training a data-driven model,it is often assumed that the data distribution of the source domain and the target domain are the same.However,in the natural scenario,this assumption is usually not tenable,and it is easy to cause poor generalization ability of the model.Domain adaptation is a method proposed to improve the generalization ability of the model.It aligns the data distribution of the source domain and the target domain by learning the data characteristics of the two domains,so that the model trained in the source domain data can also perform well in the target domain with a small number of data labels or without data labels.In order to further improve the generalization ability of the model,existing researches have explored the know-ledge integrating into domain adaptive methods,which has high practical value.Firstly,we summarizes the development background of domain adaptive methods with knowledge integrating and the research status of related reviews.Then,the problem defi-nition and theoretical basis of domain adaptation are introduced.After that,a classification system of domain adaptive methods with knowledge integrating is presented,and some representative methods are summarized.Finally,through the analysis of the challenging problems in this field,the future research directions of domain adaptive methods with knowledge integrating are predicted,in the hope of providing some reference for related research.

Text Paraphrase Generation Based on Pre-trained Language Model and Tag Guidance

LIANG Jiayin, XIE Zhipeng

Computer Science. 2023, 50 (8): 150-156. doi:10.11896/jsjkx.221100128

Abstract

PDF(2096KB) ( 295 )

References | Related Articles | Metrics

Text paraphrase generation is an important and challenging task in NLP.Some recent works have applied the syntactic structure information of different granularity of sentences to guide the process of paraphrase generation and have achieved fair performance.However,this kind of methods are rather complex and difficult to transfer.Besides,pre-trained language model has shown good performance in various NLP tasks due to knowledge learned.But it has rarely been used in the paraphrase generation task.This paper proposes a paraphrase generation method based on pre-trained language model and tag guidance.The pre-trained language model is fine-tuned to improve the performance of the paraphrase generation task,and a simple tag insertion method is used to provide syntactic structure guidance.Experiment results show that the proposed method outperforms traditional Seq2Seq methods on datasets ParaNMT and Quora.In addition,it also demonstrate its effectiveness in improving downstream tasks by data augmentation.

Study on Enhanced Entity Representation for Document-level Relation Extraction

DING Xiaoyao, ZHOU Gang, LU Jicang, CHEN Jing

Computer Science. 2023, 50 (8): 157-162. doi:10.11896/jsjkx.220700161

Abstract

PDF(2276KB) ( 270 )

References | Related Articles | Metrics

Document-level relation extraction is a hot and challenging issue in natural language processing.Graph-based model is one of the mainstream methods of document-level relation extraction.Although this method can effectively solve the long-distance dependency between entity nodes,it often fails to fully consider the additional information such as sentence context,document topic,entity to entity distance and entity to similarity when constructing nodes,resulting in low performance of relationship extraction.A document-level relation extraction model based on enhanced entity representation is proposed to solve this problem.Firstly,the original document is used as input to construct the basic document graph structure.Then,the graph neural network propagation mechanism is used to aggregate the information of adjacent nodes,and the sentence context and topic information related to entity relation prediction is integrated into the entity node representation of the primary document graph,to obtain an enhanced entity node representation.Finally,the graph model of the enhanced entity node is used to predict the entity relationship.Experimental results show that the performance of the proposed model in the document-level relation extraction task is better than that of the existing models,and has better interpretability.

Multimodal Knowledge Graph Embedding with Text-Image Enhancement

XIAO Guiyang, WANG Lisong , JIANG Guohua

Computer Science. 2023, 50 (8): 163-169. doi:10.11896/jsjkx.220700216

Abstract

PDF(1820KB) ( 252 )

References | Related Articles | Metrics

Most traditional knowledge representation learning methods only focus on the structured information in triples,and cannot make good use of the additional information such as entity images,relation path and text descriptions to learn knowledge representation or fuse only one additional information.Therefore,a multimodal knowledge graph embedding method combining entity descriptions and images is proposed.Through mutual enhancement of text and images,more comprehensive external information can be provided to make up for the deficiency of knowledge representation learning caused by the incompleteness of a single information source.Firstly,text representation and image representation of entities are obtained by modeling entity descriptions and images.Then,they are used as a supplement to the structural representation in TransE.Finally,through the joint trai-ning of three entity representations,the unified spatial representation of knowledge graph,text and image is realized to improve the accuracy of entity and relation prediction.Experimental results show that the hit rate of entity prediction of this model improves by 3.09% compared with the method of without additional information,improves by 0.97% compared with the method of fusing only entity descriptions,and improves by 1.32% compared with the method of fusing only entity images.

Answer Extraction Method for Reading Comprehension Based on Frame Semantics and GraphStructure

YANG Zhizhuo, XU Lingling, Zhang Hu, LI Ru

Computer Science. 2023, 50 (8): 170-176. doi:10.11896/jsjkx.220600070

Abstract

PDF(2564KB) ( 212 )

References | Related Articles | Metrics

Machine reading comprehension is one of the most challenging tasks in the field of natural language processing.With the continuous development of deep learning technology and the release of large-scale MRC datasets,the performance of MRC models keep breaking records.However,the previous models still have shortcomings in logical reasoning and deep semantic understanding.In order to solve the above problems,this paper proposes a reading comprehension answer extraction method based on frame semantics and graph structure.This method first uses Chinese FrameNet to match candidate sentences related to question semantics.Secondly,the entities in the questions and candidate sentences are extracted,and the entity relationship graph is constructed based on the dependent syntax and semantic relationships of entities in the sentences.Finally,the entity relationship graph is introduced into the graph attention network for logical reasoning,so as to realize the extraction of reading comprehension answers.Experiment results on Dureader-robust dataset show that the proposed method achieves better results than the baseline model.

Link Prediction Model on Temporal Knowledge Graph Based on Bidirectionally Aggregating Neighborhoods and Global Aware

TANG Shaosai, SHEN Derong, KOU Yue, NIE Tiezheng

Computer Science. 2023, 50 (8): 177-183. doi:10.11896/jsjkx.220900061

Abstract

PDF(2154KB) ( 240 )

References | Related Articles | Metrics

Temporal knowledge graphs(TKG) have great potential of application in many fields,such as recommender systems,search engine and natural language processing,but the incompleteness of TKG limites its application,so it is important to study link prediction model on TKG.Most existing methods focus on TKG completion and can't predict future facts.This paper proposes a link prediction model on TKG,which is based on bidirectionally aggregating neighborhoods and global aware.On the one hand,the proposed model independently aggregates entity's recently active and positive behavior and models their temporal evolution by recurrent neural network (RNN).On the other hand,it captures the chronic behavior patterns of entities by global aware module.Experimental results on four benchmark datasets show that our proposed method can improve the performance of forecasting future facts.

Single-stage Joint Entity and Relation Extraction Method Based on Enhanced Sequence Annotation Strategy

ZHU Xiubao, ZHOU Gang, CHEN Jing, LU Jicang, XIANG Yixin

Computer Science. 2023, 50 (8): 184-192. doi:10.11896/jsjkx.220700082

Abstract

PDF(2807KB) ( 288 )

References | Related Articles | Metrics

Extracting entities and relations from unstructured text is the fundamental task of automatically constructing know-ledge bases.Existing works mainly adopt joint learning to solve the problems of nested entities,overlapping relations,redundant computation,or exposure bias,but a single model only performs well on some issues,and no model can solve the above problems simultaneously.Therefore,a single-stage joint entity and relation extraction method based on an enhanced sequence annotation strategy called ATMREL is proposed.First,an enhanced sequence annotation strategy is designed to tag each word in the text with multiple labels,and the labels contain information about the position of each word in the entity,the relation type and the entity location.Second,the labels prediction of each word is transformed into a multi-label classification task,while the joint entity and relation extraction is transformed into a sequence annotation task.Finally,to enhance the dependencies between entity pairs,an entity correlation matrix is introduced for pruning the extraction results to improve the model extraction effect.Experimental results show that ATMREL model reduces the parameter volume by 3.1×10⁶~5.4×10⁶,improves the average inference speed by 2~4.2 times,and improves the F1 value by 0.5%~2.1% compared with the CasRel and TPLinker models on the NYT and WebNLG datasets.

Human Activity Recognition with Meta-learning and Attention

WANG Jiahao, ZHONG Xin, LI Wenxiong, ZHAO Dexin

Computer Science. 2023, 50 (8): 193-201. doi:10.11896/jsjkx.220900124

Abstract

PDF(3797KB) ( 348 )

References | Related Articles | Metrics

With the in-depth research of deep learning technology,its application and development in the field of behavior recognition have been greatly promoted.Current research on behavior recognition based on deep learning usually requires a large training data set.But when facing practical applications,new users will inevitably run into personalization issues.This means that even while performing the same activity,different people may use training data sets differently.Existing solutions cannot guarantee to achieve the expected accuracy when dealing with new users.Besides,these models would also be impractical to deploy when gathe-ring training data for new users.Facing this problem,small-sample learning can achieve better results by using only a small number of samples.This means that in the behavior recognition problem,each new user can be classified using a little training data.In this paper,a MAML-M model is proposed by combining few-shot learning and behavior recognition algorithms.Firstly,an optimization-based meta-learning method is adopted to divide the dataset according to users and construct multiple user tasks for trai-ning and testing.Meanwhile,the MAML method and the memory module based on the attention mechanism are introduced into the MAML-M model,which finally improves the ability of the model network to extract and summarize data features.Through experiment on MEx dataset,the proposed MAML-M model shows better performances under small sample sets.

Value Factorization Method Based on State Estimation

XIONG Liqin, CAO Lei, CHEN Xiliang, LAI Jun

Computer Science. 2023, 50 (8): 202-208. doi:10.11896/jsjkx.220500270

Abstract

PDF(2393KB) ( 213 )

References | Related Articles | Metrics

Value factorization is a popular method to solve cooperative multi-agent deep reinforcement learning problems,which factorizes joint value function into individual value functions according to IGM principle.In this method,agents select actions only according to individual value functions based on local observation,which leads to agents cannot effectively use global information to learn strategy.Although many value factorization algorithms extract the features of global state to weight individual value functions by many approaches,including attention mechanism,super network,and et al,so as to indirectly utilize global information to train agents,but this utilization is pretty limited.In a complex environment,it is difficult for agents to learn effective stra-tegies and their learning efficiency is poor.In order to improve agents' policy learning ability,an optimized value factorization method based on state estimation(SE-VF) is put forward,which introduces a state network to extract the features of global state and get a state value,and then take state loss value as part of the loss function to update agents network parameters,so as to optimize the strategy selection process of agents.Experimental results show that SE-VF performs better than QMIX and other baselines in multiple scenarios of the StarCraft 2 micromanagement mission test platform.

Chaos COOT Bird Algorithm Based on Cauchy Mutation and Differential Evolution

ZHOU Xuequan, DU Nisuo, OUYANG Zhi

Computer Science. 2023, 50 (8): 209-220. doi:10.11896/jsjkx.220500275

Abstract

PDF(4578KB) ( 245 )

References | Related Articles | Metrics

Aiming at the problems of low optimization accuracy,easy to fall into local optimization and slow convergence speed of COOT bird algorithm,a logistic chaos CDLCOOT algorithm based on Cauchy mutation and differential evolution is proposed.Firstly,the position of the COOT bird is disturbed by Cauchy mutation to expand the search range and improve the global search ability of the algorithm.Secondly,the differential evolution strategy is adopted for the leader COOT bird to increase the population diversity,so that the leader with better fitness can lead the population to search for the optimal solution,guide the individual COOT bird to move towards the optimal solution,and help it search faster.Finally,the logistic chaos factor is added to the chain movement of the COOT bird,so as to realize the chaotic chain following movement and improve the ability of the algorithm to jump out of the local optimum.Simulation experiments are carried out on 12 classical test functions and 9 CEC2017 test functions.The CDLCOOT algorithm is compared with other advanced algorithms,such as the sine cosine algorithm(SCA),gray wolf optimizer(GWO),ant lion optimizer(ALO),multi-verse optimizer(MVO),as well as original COOT bird algorithm and the original algorithm with single strategy to verify the effectiveness of the improved algorithm.Experimental results show that CDLCOOT has better global optimization ability and faster convergence speed than other heuristic algorithms and improved algorithms.In the classical test functions,the average value of the algorithm is 76 orders of magnitude higher than that of the original algorithm on the four unimodal functions.The theoretical optimal value is found on two multimodal functions,and the average value on other two multimodal functions is 3 or 4 orders of magnitude higher than the original algorithm.On the four fixed dimension multimodal functions,the algorithm can find the theoretical optimal value,and the convergence speed is faster.In CEC2017 test functions,the optimization accuracy of the algorithm in unimodal,multimodal and hybrid functions is improved compared with the original algorithm,and its convergence speed is also superior to the original algorithm and other algorithms,and the stability of the algorithm is better.

Study and Evaluation of Spiking Neural Network Model Based on Bee Colony Optimization

MA Weiwei, ZHENG Qinhong, LIU Shanshan

Computer Science. 2023, 50 (8): 221-225. doi:10.11896/jsjkx.220700181

Abstract

PDF(2820KB) ( 233 )

References | Related Articles | Metrics

In order to improve the training ability of Spiking neural network,this paper takes multi-label classification problem as the research breakthrough point and adopts bee colony algorithm to optimize the model.There are many neural network models based on the concept of Spiking.Probabilistic Spiking neural network(PSNN) is selected for multi-label classification.Firstly,a probabilistic Spiking neural network classification model is established.The ignition time sequence is coded,and the pulse res-ponse is triggered to realize data transmission.Then,the weight,dynamic threshold and forgetting parameters of Spiking neural network are used to construct bee colony,and the accuracy of multi-label classification is used as the fitness function of artificial bee colony(ABC) algorithm,so that the optimal individual can be obtained by constantly updating the fitness value of individual bee colony.Finally,the multi-label classification of probabilistic Spiking neural network is completed with the optimal parameters.Experimental results show that ABC-PSNN algorithm can achieve high multi-label classification accuracy by reasonably setting the individual size of bee colony and honey source search range.Compared with other Spiking neural network models and commonly used multi-label classification algorithms,ABC-PSNN algorithm has higher classification accuracy and stability.

Fusion Neural Network-based Method for Predicting LncRNA-disease Association

LI Qiaojun, ZHANG Wen, YANG Wei

Computer Science. 2023, 50 (8): 226-232. doi:10.11896/jsjkx.221000202

Abstract

PDF(3418KB) ( 239 )

References | Related Articles | Metrics

Aberrant expression of long non-coding RNA(LncRNA) is closely associated with the physiological and pathological processes of diseases.Identifying potential associations between LncRNA and diseases are helpful to understand the molecular pathogenesis of diseases.Previous researches were scarcely integrated with heterogeneous multi-source data and seldom learned high-dimensional feature representations.In this paper,we propose a new method named FNNLDA,which based on fusion neural networks(FNN) to predict the associated LncRNAs of candidate disease.FNNLDA integrates multiple data related to LncRNAs,diseases,and miRNAs.And employs the idea of multi-model fusion to learn high-level features of LncRNA-disease pairs by using two deep learning models:stacked self-encoder and fusion neural network,separately.Finally,fusing the prediction scores of the two modules to predict the LncRNA-disease associations.Five-fold cross-validation test show that the AUC value of FNNLDA method is 12.5%,15.1%,3.4% and 5.8% higher than that of SIMCLDA,MFLDA,CNNLDA and LRLSLDA,respectively.It indicates that this method has a significant improvement in LncRNA-disease prediction.The results of the study based on stomach cancer disease cases demonstrate that FNNLDA can effectively identify potential LncRNAs associated with disease.

Edge Offloading Framework for D2D-MEC Networks Based on Deep Reinforcement Learningand Wireless Charging Technology

ZHANG Naixin, CHEN Xiaorui, LI An, YANG Leyao, WU Huaming

Computer Science. 2023, 50 (8): 233-242. doi:10.11896/jsjkx.220900181

Abstract

PDF(2442KB) ( 440 )

References | Related Articles | Metrics

A large amount of underutilized computing resources in IoT devices is what mobile edge computing requires.An edge offloading framework based on device-to-device communication technology and wireless charging technology can maximize the utilization of computing resources of idle IoT devices and improve user experience.The D2D-MEC network model of IoT devices can be established on this basis.In this model,the device chooses to offload multiple tasks to multiple edge devices according to the current environment information and the estimated device state.It applies wireless charging technology to increase the success rate of transmission and computation stability.The reinforcement learning method is used to solve the joint optimization allocation problem,which aims to minimize the computation delay,energy consumption,and task dropping loss as well as maximize the utilization of edge devices and the proportion of task offloading.In addition,to adapt to larger state space and improve learning speed,an offloading scheme based on deep reinforcement learning is proposed.Based on the above theory and model,the optimal solution and upper limit of performance of the D2D-MEC system are calculated by mathematical derivation.Simulation results show that the D2D-MEC offloading model and its offloading strategy have better all-around performance and can make full use of the computing resources of IoT devices.

Analysis and Prediction of Cloud VM CPU Load Based on EMPC-BCGRU

XIE Tonglei, DENG Li, YOU Wenlong, LI Ruilong

Computer Science. 2023, 50 (8): 243-250. doi:10.11896/jsjkx.220600264

Abstract

PDF(3589KB) ( 281 )

References | Related Articles | Metrics

Cloud platform resource prediction is of great significance for resource management and energy saving.Cloud VM technology is a virtualization method implemented by the cloud to make full use of physical resources,but effective cloud VM load prediction is still challenging,because the cloud VM load has periodic and aperiodic change patterns and sudden load peaks,and the cloud VM load is affected by the random submission of jobs by users.In order to accurately analyze the change mode of VM load and improve the performance of VM CPU load prediction,a cloud VM load prediction method based on decomposition-prediction is proposed.Through EMD and PCA of cloud VM load mode decomposition,the characteristic fluctuation sequences of different time scales are obtained.The convolution layer of the prediction model can fully extract the decomposed features,and learn the forward and backward dependencies of the sequence through the bidirectional gated cyclic neural network,which improves the ability of the prediction model to learn the load change mode of the VM.Finally,single-step and multi-step prediction experiments are performed on the 2019 VM data sets generated by Microsoft Azure in the real cloud environment,which verifies the effectiveness of the prediction method.

Survey of DGA Domain Name Detection Based on Character Feature

WANG Yu, WANG Zuchao, PAN Rui

Computer Science. 2023, 50 (8): 251-259. doi:10.11896/jsjkx.220700277

Abstract

PDF(1623KB) ( 290 )

References | Related Articles | Metrics

Recent years have seen extensive adoption of domain generation algorithms(DGA) by botnets.Efficient detection of DGA domain name is of great significance for discovering botnets and ensuring cyber security.DGA domain name detection me-thod based on character feature can complete the detection only by using the domain name string.It is a real-time detection me-thod,and has become a hot spot in the research on DGA domain name detection.Research on such methods shows DGA domain name can be effectively detected by using traditional machine learning or deep learning.However,for wordlist-based DGA domain name,shorter-length DGA domain name,or new variant DGA domain name,it is still necessary to improve the detection ability by improving word embedding method,introducing attention mechanisms,or joining adversarial samples,etc.Finally,this paper summarizes the above methods,analyzes their advantages and existing problems,and proposes future research directions and key issues that need to be addressed for DGA domain name detection.

Attack Economics Based Fraud Detection for MVNO

LI Yang, LI Zhenhua, XIN Xianlong

Computer Science. 2023, 50 (8): 260-270. doi:10.11896/jsjkx.221000103

Abstract

PDF(3070KB) ( 243 )

References | Related Articles | Metrics

Driven by the full utilization of telecommunication resources and stimulating healthy market competition,mobile virtual network opera-tors(MVNOs) become popular rapidly in recent years.MVNOs rely on the infrastructures of mobile network ope-rators(MNOs) to provide users with cheaper and more flexible services.Due to the high maintenance costs of physical stores,MVNOs mostly provide fully online service.However,scammers use vulnerabilities in online authentication to purchase SIM cards and make scam calls,which seriously affects the reputation of MVNOs and their users.This has become a bottleneck problem for the survival and development of MVNOs.To address this issue,we collaborate with a large commercial MVNO with over 2 million users named Xiaomi Mobile.Related work generally assumes that scam calls are random,scattered or hidden,ma-king the detection methods inefficient or even invalid for the scenario of MVNOs.However,by analyzing the crowdsourced dataset,almost all scam calls are found to be organized,planned,and scaled.Thus,a method based on attack economics and reasonable analysis of the spatio-temporal characteristics of scam calls is proposed.This method successfully extracts the key features,and by combining with machine learning-based classification,it greatly reduces the proportion of scammers in Xiaomi Mobile to 0.023‰,which is far lower than the 0.1‰ achieved by the MNOs that have sufficient information.Under the premise of excluding the risk of being cracked,part of the code and data has been open sourced to help purify the ecology of entire telecom industry.

Spacecraft Rendezvous Guidance Method Based on Safe Reinforcement Learning

XING Linquan, XIAO Yingmin, YANG Zhibin, WEI Zhengmin, ZHOU Yong, GAO Saijun

Computer Science. 2023, 50 (8): 271-279. doi:10.11896/jsjkx.220700210

Abstract

PDF(2872KB) ( 288 )

References | Related Articles | Metrics

With the increasing complexity of spacecraft rendezvous and docking tasks,the requirements for its efficiency,autonomy and reliability are highly demanded.In recent years,the introduction of reinforcement learning technology to solve the problem of spacecraft rendezvous and guidance has become an international frontier hotspot.Obstacle avoidance is critical for safe spacecraft rendezvous,and the general reinforcement learning algorithm does not impose safety restrictions on space exploration,which make the design of spacecraft rendezvous guidance policy challenging.This paper proposes a spacecraft rendezvous guidance method based on safe reinforcement learning.First,a Markov model of autonomous spacecraft rendezvous in collision avoidance scenarios is designed,a reward mechanism based on obstacle warning and collision avoidance restraint is proposed,and thus a safe reinforcement learning framework for solving spacecraft rendezvous guidance strategy is established.Second,with the framework of safe reinforcement learning,guidance policies are generated based on two deep reinforcement learning algorithms,proximal po-licy optimization(PPO) and deep deterministic policy gradient(DDPG).Experimental results show that the method can effectively avoid obstacle and complete the rendezvous with high accuracy.In addition,the performance and generalization ability of the two algorithms are analyzed,which proves the effectiveness of the proposed method.

Facial Physical Adversarial Example Performance Prediction Algorithm Based on Multi-modal Feature Fusion

ZHOU Fengfan, LING Hefei, ZHANG Jinyuan, XIA Ziwei, SHI Yuxuan, LI Ping

Computer Science. 2023, 50 (8): 280-285. doi:10.11896/jsjkx.221100124

Abstract

PDF(2208KB) ( 287 )

References | Related Articles | Metrics

Facial physical adversarial attack(FPAA) refers to a method that an attacker pasting or wearing physical adversary examples,such as printed glasses,paper,to make the face recognition system to recognize his face as the face of a specific target,or make the face recognition system unable to recognize his face under the camera.The existing performance evaluation process of the FPAA can be affected by multiple environmental factors and require multiple manual operations,resulting in very low efficiency of performance evaluation.In order to reduce the workload of evaluating the performance of facial physical adversarial examples,combined with the multimodality between digital images and environmental factors,a multimodal feature fusion prediction algorithm(MFFP) is proposed.Specifically,different networks are used to extract the features of attacker's face images,victim's face images and facial digital adversarialexample images,and the proposed environmental feature extraction network is used to extract the features of environmental factors.A multimodal feature fusion network is proposed to fuse these features.The output of the multimodal feature fusion network is the cosine similarity performance between the predicted facial physical adversarial example image and the victim image.MFFP algorithm achieves a regression mean square error of 0.003 under the experimental scenario of unknown environment and unknown FPAA,which is better than the performance of the baseline.It verifies the accuracy of MFFP algorithm for predicting of the performance of FPAA.Moreover,it verifies that MFFP can quickly evaluate the performance of FPAA,while greatly reduce the workload of manual operation.

Study on Optimized Offloading for Data Security in Industrial Scene

WANG Biao, WANG Da, KE Ji, MA Yuqing, ZHANG Yipu, WANG Changqing, LI Aijun

Computer Science. 2023, 50 (8): 286-293. doi:10.11896/jsjkx.230100082

Abstract

PDF(2840KB) ( 237 )

References | Related Articles | Metrics

The problem of security offloading in data transmission in industrial scenarios has gained wide attention.This paper is the first to integrate security policy as a decision variable into the optimization problem.It applies computational offloading principles and differential evolutionary algorithms,and proposes a data security offloading algorithm.Firstly,mathematical modeling conducted for four computing modes of industrial field devices:local computing,local edge computing,cross-plant edge computing in this paper,and cloud computing,as well as data security,and a data security offloading model is constructed by integrating multi-level security policies,task offloading,and resource allocation.Then,the security-optimized offloading scheme is formed by designing the objective function of maximizing device satisfaction by considering the effects of time delay and security risk probability.Finally,for this optimization problem,a data security offloading algorithm based on an improved differential evolution stra-tegy is proposed to maximize the device satisfaction of the system while satisfying the optimal solution with the latency and secu-rity risk requirements.Compared with the GASORA,GSOJRA and DEDSTO-NS algorithms,the proposed algorithm enables the field devices to satisfy the delay and risk probability requirements.Furthermore,it improves the device satisfaction by 35% while guaranteeing data security.Simulation results confirm the effectiveness of the proposed method and have some realistic application value.

Reversible Data Hiding Scheme in NTRU Encrypted Domain Based on Polynomial Partition

LIU Dingcai, WU Haotian, ZHUANG Zhenwei, HE Junhui

Computer Science. 2023, 50 (8): 294-303. doi:10.11896/jsjkx.220800245

Abstract

PDF(3131KB) ( 213 )

References | Related Articles | Metrics

With the rapid development of cloud computing techniques and demand of privacy preservation,reversible data hiding(RDH) in homomorphic encrypted domain has become a hot research topic.Most of the existing RDH schemes in encrypted domain exploit correlations between adjacent pixels and redundancy in images,whose applications are limited.To improve applicabi-lity and embedding capacity,a new RDH scheme in NTRU encrypted domain based on polynomial partitioning is proposed.It divides the polynomial space in NTRU cryptosystem,which can be applied to multiple encrypted media content for data hiding.Part of the space is used to represent the original plaintext,while the rest space is used to hide the hidden data.The receiver can retrieve part of the hidden data directly from the ciphertext,while the rest hidden data can be extracted after decryption and the original plaintext can be correctly restored.In our experiments,grayscale images and text files are chosen to verify feasibility of the proposed scheme.Experimental results show that a maximum of N-8 bits can be hidden into a ciphertext for a plaintext represented with 8 bits,where N is a parameter used in NTRU cryptosystem.When N is set to 503,at most 495 bits can be hidden in a ciphertext while the plaintext can be exactly recovered.Compared with the existing schemes,the proposed scheme has higher embedding capacity and better applicability.

Two-layer IoT Device Classification Recognition Model Based on Traffic and Text Fingerprints

ZHU Boyu, CHEN Xiao, SHA Letian, XIAO Fu

Computer Science. 2023, 50 (8): 304-313. doi:10.11896/jsjkx.220900145

Abstract

PDF(4682KB) ( 260 )

References | Related Articles | Metrics

In order to isolate the vulnerable and abnormal IoT devices in the local area network in time,efficient device classification and identification capability is very important for network administrators.The features selected in the existing methods are not highly correlated with equipment,and the sample data is unbalanced due to differences in equipment status.Aiming at the above problems,this paper proposes an IoT device classification and identification model FT-DRF based on traffic and text fingerprints.This method firstly designs a feature mining model,selects stable flow statistics as device traffic fingerprints,and then generates device text fingerprints based on sensitive text information in the header fields of application layer protocols such as HTTP,DNS,and DHCP.On this basis,the data is preprocessed and the feature vector is generated.Finally,a machine learning algorithm based on double-layer random forest is designed to classify and identify the devices.A supervised classification and re-cognition experiment is conducted on the simulated smart home environment dataset composed of 13 IoT devices and public dataset.The results show that the FT-DRF model can identify IoT devices such as network cameras and smart speakers,with an ave-rage accuracy rate of 99.81%,which is 2%~5% higher than that of the existing typical methods.

Compiler-supported Program Stack Space Layout Runtime Randomization Method

ZHU Pengzhe, YAO Yuan, LIU Zijing, XI Ruicheng

Computer Science. 2023, 50 (8): 314-320. doi:10.11896/jsjkx.220800098

Abstract

PDF(2291KB) ( 209 )

References | Related Articles | Metrics

Multi-variant execution is one of the most popular active defense technologies.MVX identifies attack behavior by running a set of functionally equivalent heterogeneous variants parallelly and detecting inconsistent state transitions between different variants.The defense effect of MVX depends on the heterogeneity between program variants in a large extent.Generally,the higher the heterogeneity between program variants,the better the defense effect of MVX.To improve the heterogeneity between program variants,this paper proposes a compiler-supported,dynamic and static program stack space layout randomization me-thod.The method is based on LLVM 12.0 compilation framework.At static compile stage,the method identifies the key variables in program based on external input acquisition functions,locates their stack space allocation instructions,and adds additional call and allocation instructions before these allocation instructions.At program runtime,the method uses the instructions added during static compilation to randomly fill memory blocks before the key variables in stack space,realizing program memory space layout runtime randomization.Simulation experiment results indicate that the dynamic and static program stack space layout randomization method proposed in this paper can effectively improve the heterogeneity between MVX programs.For attacks based on program memory address overflow,the method not only increases their own attack difficulty,but also makes it impossible to conduct attacks by constantly testing program addresses,improving the defense ability of program effectively.

Privacy-preserving Data Classification Protocol Based on Homomorphic Encryption

LU Xingyuan, CHEN Jingwei, FENG Yong, WU Wenyuan

Computer Science. 2023, 50 (8): 321-332. doi:10.11896/jsjkx.220700130

Abstract

PDF(2140KB) ( 310 )

References | Related Articles | Metrics

With the development of big data and cloud computing,the demand for cloud computing services is growing dramatically.When users apply for cloud computing services,their privacy data needs to be stored and computed on cloud platforms,which may cause leakage of private data.Homomorphic encryption allows direct computation on ciphertexts,and the decryption of the resulting ciphertext is the same as computing on plaintexts,so homomorphic encryption can protect users' private data.Here a framework for two parties in the semi-honest model is considered.The client encrypts the privacy data into ciphertext according to a homomorphic encryption scheme and sends it to the server,and the server uses the plain machine learning model to classify the encrypted data from the client.Finally,the server sends the encrypted classification result back to the client,and the client decrypts the classification result by itself.With the framework above,three machine learning classifiers,the hyperplane,decision tree,and k-nearest neighbor classifier,based on the Brakerski-Gentry-Vaikuntanathan(BGV) homomorphic encryption scheme are investigated.According to the characteristics of each classifier,different ciphertext data packaging strategies and calculation processes are designed with single-instruction-multiple-data (SIMD) technology,which significantly reduces the communication overhead between the client and the server.In the prediction phase,the hyperplane and decision tree classifiers achieve interaction-free,and the KNN classifier only needs one interaction.Moreover,the three classifiers are implemented with a homomorphic encryption library HElib.For several UCI public datasets,the hyperplane classifier can complete the privacy-preserving classification within tens of milliseconds to hundreds of milliseconds for a single sample,and the decision tree can complete it within tens of milliseconds.The prediction accuracy of the first two classifiers for ciphertext data exceeds 90%,and the two parties only need the communication cost of the client sending the encrypted private data to the server,and the server returns the encrypted classification label to the client.The k-nearest-neighbor classifier completes one sample's classification in about 4 seconds on average,and the prediction accuracy of ciphertext data is also more than 90%.In addition to the communication overhead of privacy data and classification labels,the two parties also need an additional round of intermediate calculation results between the server and the client to complete the classification.Compared with similar protocols based on homomorphic encryption,the proposed protocols have advantages in the number of communication rounds,prediction accuracy,and computational efficiency.

Recoverable Data Aggregation Protocol for Wireless Sensor Networks Based on Reversible DigitalWatermarking

GAO Guangyong, HAN Tingting, XIA Zhihua

Computer Science. 2023, 50 (8): 333-341. doi:10.11896/jsjkx.220800089

Abstract

PDF(2294KB) ( 214 )

References | Related Articles | Metrics

Aiming at the opposition between the high energy consumption of data security authentication protocol and the resource limitation of sensor nodes in wireless sensor networks,this paper proposes an aggregation protocol based on reversible di-gital watermarking.On the one hand,at the sensing node,the watermarking is embedded into the sensing data,and the elliptic curve is used to encrypt the watermarked data homomorphically,so as to ensure the privacy of the data in the transmission process.At the cluster head node,the received data are only performed aggregation and forwarding operations,so as to reduce network communication overhead.At the base station,the watermark is extracted to authenticate the integrity of data.On the other hand,this scheme proposes an aggregation tree protocol based on clustering protocol,which can reduce the transmission energy consumption of nodes and prolong the network lifetime.Theoretical analysis proves that the proposed protocol combines watermarking technology and data aggregation technology better,has good security and lower computation cost,and can realize the integrity authentication of perception data.In addition,experimental results show that,compared with the latest similar algorithms,the proposed protocol has certain advantages in communication cost and delay.

Differential Privacy Linear Regression Algorithm Based on Principal Component Analysis andFunctional Mechanism

LI Kejia, HU Xuexian, CHEN Yue, YANG Hongjian, XU Yang, LIU Yang

Computer Science. 2023, 50 (8): 342-351. doi:10.11896/jsjkx.220800255

Abstract

PDF(2185KB) ( 298 )

References | Related Articles | Metrics

With the continuous development of artificial intelligence applications and the subsequent promulgation of privacy protection laws and regulations,the privacy protection issue in machine learning has become a research hotspot in the field of information security.To overcome the issues of high global sensitivity and poor model usability of the existing differential privacy linear regression algorithms,we present a differential privacy linear regression algorithm based on principal component analysis and functional mechanism(PCAFM-DPLR).In the PCAFM-DPLR algorithm,the traditional Laplace mechanism is replaced by the Gaussian mechanism,and the noise is added in the two major stages of the algorithm respectively.First,in order to take into account the privacy of the data while reducing the dimensionality,Gaussian noise is injected into the covariance matrix of the original data set,and a low-dimensional data set with differential privacy protection effect is obtained based on principal component analysis.Second,to prevent the possible privacy leakage during the model training,Gaussian noise is then added to the expansion polynomial coefficients of the objective function,and the minimization of the perturbed objective function is used as the objective to find the optimal model parameters.Theoretical analysis and experimental results show that the linear regression model trained by the PCAFM-DPLR algorithm can effectively guarantee privacy while having good utility.

Opaque Predicate Construction Algorithm Without Size Constraints

WANG Yufang, LE Deguang, Jack TAN, XIAO Le, GONG Shengrong

Computer Science. 2023, 50 (8): 352-358. doi:10.11896/jsjkx.220600149

Abstract

PDF(2230KB) ( 217 )

References | Related Articles | Metrics

Combined with opaque predicate,control flow obfuscation enables semantics-preserving transformations,which can achieve the purpose of code protection.However,existing opaque predicate is easily attacked by symbolic execution and has the problem of small symbolic variable.To solve the above problems,combined with symbolic variable and array,this paper designs the conditional expression of inequality by single array nesting and modulo add operation of symbolic variable,based on which an algorithm for constructing opaque predicate without size constraints is proposed.The opaque predicate obfuscation based on the proposed algorithm can incur not only false negative but also false positive issues to attackers,which effectively defends against symbolic execution attacks.Besides,the potency,resilience and cost of the program obfuscated by opaque predicate without size constraints are experimentally tested and analyzed by measuring procedures such as opaque predicate detection,bogus control flow removal and so on.Experimental results show that the opaque predicate obfuscation based on the proposed algorithm not only demonstrates excellent potency and efficient cost,but also has high resilience to anti-deobfuscation in new test environment.

Anonymous Authentication Protocol for Medical Internet of Things

LIU Yingjun, LUO Yang, YANG Yujun, LIU Yuanni

Computer Science. 2023, 50 (8): 359-364. doi:10.11896/jsjkx.220700151

Abstract

PDF(1632KB) ( 238 )

References | Related Articles | Metrics

As IoT technology continues to mature,it has been frequently used in various industries to improve people's work efficiency and living standards.The widespread application of IoT in the medical field facilitates patients' access to medical services while also allowing doctors to obtain more timely and accurate information about the patient's physical condition,so that they can develop more efficient treatment plans.However,while people are enjoying the convenience of medical IoT,how to ensure the communication security and personal privacy of patients are issues that cannot be ignored.In order to realize users' secure access to the network,this paper proposes an efficient anonymous authentication and key exchange protocol based on homomorphic encryption.Medical devices and telemedicine servers only need a low-entropy password for mutual authentication,thus negotiating a high-entropy session key.In this paper,the security of the scheme is proved under the standard model,and the simulation experimental results show that the scheme is more efficient than existing similar schemes.

Blockchain-based Dual-branch Structure Expansion Model

WANG Junlu, LIU Qiang, ZHANG Ran, JI Wanting, SONG Baoyan

Computer Science. 2023, 50 (8): 365-371. doi:10.11896/jsjkx.220900049

Abstract

PDF(3069KB) ( 235 )

References | Related Articles | Metrics

With the rapid development of blockchain technology,blockchain faces scalability challenges in terms of storage overhead and data throughput.The blockchain is affected by the consensus principle of overall consensus,and the global ledger of the entire blockchain needs to be stored between nodes,and the data storage overhead is high.At the same time,in order to maintain the consistency and credibility of transactions within the block,all nodes participate in the process of transaction verification and synchronization,the block synchronization delay in the peer-to-peer network is high.And the bandwidth requisition is blocked,which further reduces the data throughput.In response to these problems,this paper proposes a blockchain-based dual-branch structure expansion model.First,a ternary storage expansion structure of the blockchain is established.The nodes accurately divide the storage tasks and store the single,partial,and global ledger of the blockchain,which effectively reduces the storage burden of the nodes.Secondly,a dual-branch structure model is proposed,the main chain is divided into multi-channel sub-chains.And data is stored in parallel through multi-channel sub-chains,which significantly improves the data storage rate.Aiming at the compatibility problem of sub-chains after shunting,a two-way rotation mechanism is introduced to realize the fusion transition between chain structures.For the security problem of sub-chains after shunting,the gambler extension-F and gambler extension-S strategies are proposed to simulate the security attack of the two chain structures,and the mathematical modeling of the attack process is carried out.Finally,constructing the security constraints of the two models to verify the security of the dual-branch model.Experiments show that the dual-branch structure expansion model proposed in this paper can effectively resist malicious double-spending attacks,and has great advantages in storage overhead and data throughput.