Research And Application Of Multi Task Bug Fixing Based On Deep Learning

In recent years,the number and severity of network attacks have increased rapidly.Usu-ally,attackers directly invade remote computers or use malicious software by taking advantage of existing bugs in hardware,software and network,resulting in serious hidden dangers to Cy-berspace Security.The system suffers from external exploitation attacks due to its own bugs,resulting in serious unsafe consequences such as privacy disclosure,illegal ultra vires,black-mail and so on.It is an important focus in the field of Cyberspace Security to detect the security bugs in the system as soon as possible and allocate appropriate developers in time,give priority to fixing and strengthening the bugs with high severity,effectively avoid the problems of data loss and leakage during the operation of the system,prevent virus intrusion,avoid serious losses to users,and ensure the security,privacy,integrity,practicability and controllability of informa-tion.Therefore,this paper mainly studies the two important contents of allocating appropriate developers to fix bugs and predicting the severity of bugs in the process of bug fixing.However,the current methods mainly have the following three problems: 1)the extracted text structure features can not be combined in a more effective way,resulting in poor recom-mendation effect and can not be further expanded.2)The existing methods deal with the two tasks of bug fixing recommendation and bug severity prediction respectively.Due to repetitive work,the completion of all tasks is very time-consuming.It also ignores the relationship be-tween the two tasks,and can not make full use of the useful information between tasks to assist the prediction of each subtask.3)Because the category labels of bug severity in the training data set are seriously unbalanced,the prediction effect of bug severity is poor.Aiming at problem 1,this paper proposes a novel vulnerability repairer recommendation algorithm RCNNA.The proposed algorithm uses CNN convolution kernel to capture the local information of the text,and RNN is used to capture the sequence information of the text.At the same time,attention mechanism is introduced into RNN to learn the contribution proportion of each part of the text to the overall semantic information of the text.The algorithm can effectively learn the vulnerability reporting feature,and recommend appropriate developers to repair the vulnerability according to the feature.On the two open source data sets of Eclipse and Mozilla,the accuracy of top5 reaches 87.36 % and 78.64 % respectively,which are better than the five baseline methods.At the same time,the impact of vulnerability report metadata information and developer activity on recommendation accuracy is analyzed.Aiming at problem 2,this paper proposes a multi task learning model RCNNA＿M.By sharing the representations between related tasks in the task sharing layer,the task specific layer learns the unique characteristics of their respective tasks,and can recommend and predict the bug fixers and the severity of bugs at the same time.Thus,the task execution time is reduced and the generalization ability of the model is improved to a certain extent.Aiming at problem 3,because the category labels of bug severity in the training data set are seriously unbalanced,the context data enhancement method is introduced through the GAN generative countermeasure network to generate bug fixing reports to balance the category labels,so as to improve the prediction effect of bug severity.