| With the continuous development of information technology,the security of computer software has been widely concerned,among which vulnerability is one of the main causes of software security threats.Nowadays,with the popularity of multi-core operating system,multithreaded programs in the execution process due to the competition of shared resources and resulting in concurrency vulnerabilities,once concurrency vulnerabilities are used,may lead to such as program crash,denial of service attacks,data damage and other results,seriously affecting the availability,integrity and security of the system,Therefore,the study of concurrency vulnerability is of great significance to solve network security problems.As a popular technical question-and-answer community,Stack Overflow contains a large number of question-and-answer posts related to concurrency vulnerabilities.How to mine the needs related to concurrency vulnerabilities concerned by users from a large number of texts has become the research focus.With the development of neural networks,neural topic models have been applied in topic mining.However,existing neural topic models fail to utilize the external knowledge in corpora and make full use of the relationship between documents to infer topics.In order to help users better understand the content of concurrency vulnerability posts,this paper takes the problem posts in the field of concurrency vulnerability in Stack Overflow as the research object,improves the existing deficiencies in the current neural theme model,and tries to mine the concurrency vulnerability topics more accurately and complete the topic analysis task.In view of the above problems,the main work of this paper is as follows:1.Considering the poor semantic representation quality of words caused by the Glove static word vector in the Gaussian neural topic model encoder based on bidirectional adversarial training,this paper adopts the ALBERT pre-training model combined with dynamic contextual semantic learning to improve the representation of document vector and further improve the quality of semantic representation.The CycleGAN cycle is introduced to better guide the encoder and generator network to conduct confrontation training.The constructed data sets are compared to verify the effectiveness of the model proposed in this paper,and a more explanatory topic word generation representation is obtained.2.The traditional topic model only makes use of the interrelationship of words in a single document in the training process,and fails to make full use of other documents and word information related to the topic of the document.In this paper,a topic modeling method GATTM based on graph attention network is proposed.By connecting all documents and words as nodes,documents and words in the concurrency vulnerability corpus are represented as document relational diagrams.GATv2 network is introduced to correct the attention function and further enhance the feature vector representation of nodes.The topic distribution of the current document is identified by information aggregation of the adjacent nodes of the document.The results show that GATTM is superior to other models in topic modeling performance.3.In this study,a data set of concurrency vulnerabilities based on Stack Overflow is constructed.After data acquisition,initial tag setting,tag expansion,data preprocessing and other processes,a total of 47,683 concurrency vulnerability posts are obtained as the final experimental data set.Based on the topic model proposed in this paper,statistical analysis and topic analysis are carried out on posts related to concurrency vulnerabilities.The ranking of popularity and difficulty under each topic is discussed,and topic analysis is carried out according to the results of topic popularity and difficulty.The hope is to find the popular and difficult questions in the field of concurrency vulnerability,and provide decisions for concurrency developers and researchers to determine the focus of work.4.Based on the topic model proposed in this study,a system for topic analysis in the field of concurrency vulnerability is designed and implemented.The system is mainly divided into user registration and login,data set upload,topic modeling,topic analysis and user download management,etc.By uploading the data set of concurrency vulnerability,users can understand popular and difficult questions. |
