Search

In recent years,machine learning has not only achieved remarkable results in conventional fields such as computer vision and natural language processing,but also been widely applied to process sensitive data such as face images,financial data and medical information.Recently,researchers find that machine learning models will remember the data in their training sets,making them vulnerable to membership inference attacks,that is,the attacker can infer whether the given data exists in the training set of a specific machine learning model.The success of membership inference attacks may lead to serious individual privacy leakage.For example,the existence of a patient's medical record in a hospital's analytical training set reveals that the patient was once a patient there.The paper first introduces the basic principle of membership inference attacks,and then systematically summarizes and classifies the representative research achievements on membership inference attacks and defenses in recent years.In particular,how to attack and defend under different conditions is described in detail.Finally,by reviewing the development of membership inference attacks,this paper explores the main challenges and potential development directions of machine learning privacy protection in the future.

Recently,the fairness of the algorithm has aroused extensive discussion in the machine learning community.Given the widespread popularity of spectral clustering in modern data science,studying the algorithm fairness of spectral clustering is a crucial topic.Existing fair spectral clustering algorithms have two shortcomings:1) poor fairness performance;2) work only for single sensitive attribute.In this paper,the fair spectral clustering problem is regarded as a constrained spectral clustering problem.By solving the feasible solution set of constrained spectral clustering,an unnormalized fair spectral clustering(UFSC) method is proposed to improve fairness performance.In addition,the paper also proposes a fair clustering algorithm suitable for multiple sensitive attribute constraints.Experimental results on multiple real-world datasets demonstrate that the UFSC and MFSC are fairer than the existing fair spectral clustering algorithms.

Artificial intelligence has been integrated into all aspects of people's daily lives with the continuous development of machine learning,especially in the deep learning area.Machine learning models are deployed in various applications,enhancing the intelligence of traditional applications.However,in recent years,research has pointed out that personal data used to train machine learning models is vulnerable to the risk of privacy disclosure.Membership inference attacks(MIAs) are significant attacks against the machine learning model that threatens users' privacy.MIA aims to judge whether user data samples are used to train the target model.When the data is closely related to the individual,such as in medical,financial,and other fields,it directly interferes with the user's private information.This paper first introduces the background knowledge of membership inference attacks.Then,we classify the existing MIAs according to whether the attacker has a shadow model.We also summarize the threats of MIAs in different fields.Also,this paper points out the defense means against MIAs.The existing defense mechanisms are classified and summarized according to the strategies for preventing model overfitting,model-based compression,and disturbance.Finally,this paper analyzes the advantages and disadvantages of the current MIAs and defense mechanisms and proposes possible research directions for future MIAs.

Reproducible build is the ability to recreate binary artifacts in a predefined build environment.Due to the role of reproducible build in ensuring the security of software construction environment and improving the efficiency of software construction and distribution,many open source software repositories(such as Debian) have carried out software reproducible build practice.However,due to the lack of sufficient judgment information and the complexity and diversity of source files,it is still a time-consuming and laborious challenge to determine why software can not be built reproducibly.In order to overcome this challenge,this paper studies the classification and detection of software unreproducible build causes based on machine learning.This paper stu-dies four typical reasons for unreproducible build,namely timestamp,fileordering,randomness and locale.This method uses the word vector generated by word2vec to represent the text log,and then cooperates with the logistic regression model to learn and train the text corpus combined with the difference log and the build log,so as to realize the automatic classification of the causes of unreproducible build.In this paper,the algorithm is implemented and tested on 671 unreproducible build Debian software packa-ges.Experimental results show that our method achieves a macro average precision of 80.75% and a macro average recall of 86.07%,which are better than other commonly used machine learning algorithms.In addition,we also analyze the relevance and importance of difference log and build log.Result indicates that both of them are significant for the classification of unreproducible build causes.This method provides a reliable research basis for automatic classification of unreproducible build causes.

With the increasing globalization of football,the global player transfer market is becoming more and more prosperous.However,as the most important factor affecting player transfer transaction,the player’s transfer value lacks in-depth model and application research.In this paper,the FIFA’s official player database is taken as the research object.Firstly,on the premise of distinguishing different player positions,Box-Cox transformation,F-Score feature selection,etc.are used to perform feature processing on the original data set.Secondly,the player value prediction model is constructed by XGBoost,and compared with the main machine learning algorithms such as random forest,AdaBoost,GBDT and SVR for 10-fold cross validation experiments.Experimental results prove that the XGBoost model has a performance advantage over the existing models on the indicators of R2,MAE and RMSE.Finally,on the basis of constructing the value prediction model,this paper integrates the SHAP framework to analyze the important factors affecting the players’ value score in different positions,and provides decision support for some scenarios,such as player’s value score evaluation,comparative analysis,and training strategy formulation,etc.

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.

Federated learning is proposed to solve the contradiction between data sharing and privacy-preserving.It aims to build collaborative models by securely interacting irreversible information (e.g.,model parameters or gradient updates).However,the risks of privacy leakage and malicious attacks in the process of model local training,information interaction and parameter transmission have brought major challenges to the practical application of federated learning.This paper summarizes the Attack beha-viors and corresponding defense strategies in the modeling and deployment process of federated learning.Firstly,this paper briefly reviews the development process of federated learning and the basic modeling process.Next,it classifies attack behaviors in fede-ral learning training and deployment from three aspects:confidentiality,availability and integrity,and combs the latest research on privacy theft and malicious attacks.Then,it summarizes defense countermeasures from two directions: honest-but-curious attac-kers and malicious attackers,and analyzes the defense capabilities of different strategies.Finally,it presents some discussions about the problems and challenges of attack and defense methods in the practice of federated learning.Besides,it looks forward to their future development direction of federated learning in defense strategy and system design.

Based on blockchain technology,Ethereum Solidity smart contract as a computer protocol is designed to spread,verify,or execute contracts in an informative way,and it provides a foundation for various distributed application services.Although implemented for less than six years,its security problems have frequently broken out and caused substantial financial losses,which attracts more attention in the security inspection research.This paper firstly introduces some specific mechanisms and operating principles of smart contracts based on Ethereum related techniques,and analyzes some smart contract vulnerabilities occurring frequently and deriving from the characteristics of smart contracts.Then,this paper explains the traditional mainstream smart contract vulnerability detecting tools in terms of symbolic execution,fuzzing,formal verification,and taint analysis.In addition,in order to cope with the endless new vulnerabilities and the need to improve the efficiency of detection,vulnerabilities detection based on machine learning in recent years is classified and summarized according to the various ways of problem transformation in three perspectives including text processing,non-Euclidean graph and standard image.Finally,this paper proposes to formulate more extensive and accurate standardized information database and measurement indicators towards the insufficiency of the detection methods in two directions.

Modern compilers offer many optimization options.It is a complex problem to choose which parameter values,which combination of options and in which order to apply these options.Among them,the optimization phase ordering is the most dif-ficult one.With the improvement of traditional methods (iterative compilation combined with heuristic optimization search) and the emergence of new technologies (machine learning),it is possible to build a relatively efficient and intelligent compiler automatic tuning framework.This paper summarizes the research ideas and application methods of predecessors by investigating the related research in the past decades.Firstly,the development of compiler automatic tuning is introduced,including early manual methods,cost function driven methods,iterative compilation and machine learning-based prediction methods.Then,this paper focuses on the direct prediction based on machine learning and the automatic optimization method of iterative compilation driven by machine learning.Last but not least,several successful frameworks and some recent research results are listed.Current challenges and some key future research directions are also pointed out.

This paper proposes a user-centered location privacy game mechanism,which is aimed to generate corresponding protection strategies based on the LBS service quality,and reduce the calculation scale and utility loss.This mechanism is based on the Stackelberg game model.When a user requests a LBS service,he/she uses the location ambiguity mechanism to disturb his/her location and send it to the LBS server,making it difficult for the attacker to predict his/her real location.Based on part of their known background knowledge,attackers infer the protection policies of users in the anonymous area and adjust their attack methods to minimize the level of user privacy.In order to solve the problem of large scale and long time calculation by traditional mathematical methods,this paper adopts generating countermeasures network to participate in the generation of protection strategy,and reduces the utility cost as much as possible.The experimental results show that the protection mechanism has good performance in terms of privacy protection level,and at the same time,it significantly reduces the generation time of the protection mechanism while losing some quality of service.

In optic fiber communication,traditional optical performance monitoring(OPM) mainly relies on analyzing the time-frequency domain information of the signal.However,conventional methods cannot complete multi-task joint monitoring,so they are less flexible.With the development of machine learning,the monitoring of optical signal modulation format(MF) and optical signal-to-noise ratio(OSNR) based on machine learning have been gradually applied.However,existing methods have low accuracy for OSNR monitoring in complex scenarios because they do not consider the fine-grained characteristics of the signal.This paper proposes a joint monitoring model(FGNet) for optical signal MF and OSNR based on fine-grained constellation identification to solve this problem.Firstly,the backbone feature extraction module uses a deep residual structure.Secondly,a multilayer bilinear pooling module is proposed to perform fine-grained feature analysis on constellation features.Finally,a joint MF and OSNR monitoring module is proposed to realize the feature fusion of MF and OSNR.Extensive experiments with 7 200 constellation maps in the simulation dataset show that the proposed model has achieved superior performance compared to existing methods.

Due to the explosive growth of data in the era of big data,it is difficult for the traditional index structures to handle this huge and complex data.In order to solve this problem,the learned index has emerged and become one of the most popular research topics in the database.Learned indexes employ machine learning models for index construction.By training and learning the relationship between data and physical location,the learning model can be obtained so as to master the distribution characte-ristics between the two to realize the improvement and optimization of the traditional index.Extensive experiments show that learned indexes can adapt to large-scale data sets,and provide better search performance with lower memory requirements than traditional indexes.This paper introduces the applications of learned indexes and reviews the existing learned index models.According to data types,learned indexes are divided into two categories:one-dimensional and multi-dimensional.The advantages,disadvantages,and supported searches of learned index models in each category are also introduced and analyzed in detail.Finally,some future research directions of learned indexes are prospected to provide references for related researches.

Stencil computation is one kind of the most important loop kernels in scientific and engineering computing applications.Loop tiling can effectively improve the data locality of stencil computation and the degree of computational parallelism,but the best tile size is hard to choose.Traditional tile size selection methods usually have shortcomings in some ways of time overhead,labor cost and model accuracy.In this paper,a tile size selection method based on artificial neural network is proposed to predict the optimal tile size of three-dimensional Jacobi stencil loop programs.Experimental results show that,for 11 real stencil programs,the performance improvement of the programs using the model prediction tile size compared with the non tiling is 2% and 35% in serial and parallel tests respectively.Compared with the well-known grid search method,our method has a similar prediction accuracy,but only takes one 30 thousandth of the online time cost.In addition,compared with the Turbo-tiling method,our method improves the performance of tiled codes nearly 9% in average.

Automated guided vehicles(AGV)are crucial for the horizontal transportation of automated container terminals.Accurate estimation of the travel time of each AGV will reduce the number of idle AGV resources and increase the efficiency of the entire terminal.This paper proposes a method for travel time estimation of AGV in automated container terminals.Firstly,the target route of AGV is divided and encoded into several segments.Secondly,other routes are encoded as environment information,which depart before or after the target route.And the conflict between these routes and target route is estimated as an auxiliary task.Finally,the travel time with all encodings is calculated.The proposed method introduces the influence of path conflicts on time estimation.Experiments based on historical data of automated terminals show that,compared with static time estimation methods commonly used in AGV scenarios,the proposed method can reduce the time estimation error by more than 18%,and can estimate the travel time more accurately.

The rapid development of artificial intelligence and 5G network technology has opened a new era of interconnection of all things.The great improvement of computing power has threatened the traditional cryptographic algorithm based on the theory of computational difficulty.Data security and communication security have become key problems to be solved urgently in the era of Internet of things,hence cryptology has entered an intelligence era.The new generation of intelligent cryptology mainly consists of two core technologies:intelligent cryptographic algorithm based on neural network and intelligent cryptanalysis based on machine learning.The former uses the nonlinear characteristics of neural network to design the encryption process and improve the security of ciphertext.The latter trains the machine learning model through the clear ciphertext set to obtain the ciphertext features and improve the ciphertext decoding efficiency.This paper briefly reviews the development of cryptographic algorithms,discusses machine learning methods on intelligent cryptology,focuses on combing the latest progress of cryptographic algorithms and cryptanalysis intelligence at home and abroad,analyzes the advantages and disadvantages of intelligent cryptology at present,and discusses the research direction and challenges in the future.

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.

The key problem of financial asset allocation is asset price.Asset pricing is the core content of modern finance,which indicates that asset pricing law has always been one of the hot topics of financial research.This paper reviews the methods used by machine learning in the field of asset pricing and research progresses,classifies machine learning asset pricing method into machine learning method based on the characteristics processing and deep learning method based on end-to-end processing,compares the differences between different algorithms in principle and application scenarios,points out the applicability and limitations of the two kinds of machine learning methods,prospects the research direction on machine learning asset pricing in the future.

Socioeconomic Status (SES), an overall measure of a person's economic and social status relative to others combining factors such as economics and sociology, has received a lot of attention from researchers, as its assessment can help relevant orga-nizations to make various policies and decisions (governmental formulation of social policies, advertising personalized services, etc).In addition, with the development of big data technology and machine learning in recent years, assessing people's socioeconomic attributes (SEAs) and further obtaining the corresponding socioeconomic status with a data-driven approach can address the issue of extremely high cost of traditional methods.Therefore, this paper summarizes the research progresses of applying big data techniques to socioeconomic status analysis in recent years.It first introduces the basic concept of socioeconomic status and discusses the challenges posed by big data methods compared to traditional methods.After that, it systematically summarizes and classifies the state-of-the-art related methods based on the information in the learning process, and present them in detail, discusses the pros and cons of each type of method.Finally, it discusses the challenges and problems of inferring people's socioeconomic status and provides an outlook on future research directions.

Physical-informed neural networks (PINN) are a class of neural networks used to solve supervised learning tasks.They not only try to follow the distribution law of the training data, but also follow the physical laws described by partial diffe-rential equations.Compared with pure data-driven neural networks, PINN imposes physical information constraints during the training process, so that more generalized models can be acquired with fewer training data.In recent years, PINN has gradually become a research hotspot in the interdisciplinary field of machine learning and computational mathematics, and has obtained relatively in-depth research in both theory and application, and has made considerable progress.However, due to the unique network structure of PINN, there are some problems such as slow training or even non-convergence and low precision in practical application.On the basis of summarizing the current research of PINN, this paper explores the network/system design and its application in many fields such as fluid mechanics, and looks forward to the further research directions.

Cataract is a main ocular disease for visual impairment and blindness.Anterior segment optical coherence tomography (AS-OCT) technique has the characteristics of non-invasiveness,high resolution,rapid inspection,and objective quantitative measurement.AS-OCT images have been widely used for the diagnosis of ocular diseases in clinical ophthalmology.Inthecurrent,it is lack of the research on classification methods of nuclear cataract based on AS-OCT images.To this end,this paper proposes a nuclear cataract classification method based on AS-OCT images.First,the nucleus region of the lens is extracted from AS-OCT images using a combination of adaptive threshold method,edge detection Canny algorithm and manual correction pattern.Then,eighteen pixel features are extracted based on image intensity and histogram feature statistical methods,and the Pearson correlation coefficient method is used to analyze the correlation between the extracted pixel features and the severity of nuclear cataract.Finally,the random forest algorithm is used to build a classification model for getting nuclear cataract classification results.Experimental results on an AS-OCT image dataset show that the proposed method achieves the accuracy and recall with 75.53% and 74.04% respectively.Therefore,the proposed method has the potential as a quantitative analysis reference tool for the clinical diagnosis of nuclear cataract.

The tasks of distributed on-board multi-RTs (remote terminals) system are mainly distributed based on functions,while the burstiness of data processing tasks often leads to unbalanced load among different computers.Using a flexible load scheduling mechanism can effectively adjust the load difference between different computers,thereby improving the overall performance of the computer system to a certain extent.A load scheduling algorithm for distributed on-board RTs system based on machine learning is proposed in this paper,which includes four steps:sample collection,task throughput prediction model construction,throughput prediction and load scheduling.In the process of constructing the task throughput prediction model,the weight of the model is obtained through the linear regression normal equation of machine learning,which reduces the time spent in constructing the model.In the load scheduling link,if the total throughput rate of RTs is greater than the total load data volume of the system,data will be allocated to each RT in proportion to the throughput rate;otherwise,only a certain amount of data will be allocated to RTs whose load data volume is less than their own throughput rate.The test results on the ground simulation system constructed by multiple on-board computers electrical performance products show that the algorithm can increase the average CPU utilization rate of all nodes of the system by 23.78%,and reduce the variance of CPU utilization rate between nodes to 34.59%.The total system throughput of the task is significantly increased by 225.97%.In other words,this method can effectively improve system resource utilization while ensuring system load balance,and improve the real-time data processing performance of the on-board computer system.