Design And Implementation Of Source Code Vulnerability Detection System Based On Dynamic LSTM

As technology evolves and people’s needs increase,software development becomes more frequent and software vulnerabilities increase.To address these issues,deep learning-based source code vulnerability detection methods can automate the extraction of vulnerability features and detect vulnerabilities based on these features.In the existing research,the lack of differentiation between positive and negative samples and the low ability of neural networks to model complex code led to the problems of poor detection accuracy and low stability.In this thesis,we study a source code vulnerability detection method based on metric learning and BiPhased LSTM,whose main innovations are as follows.(1)In response to the problem that existing research does not use category comparison information and does not distinguish enough between positive and negative samples,leading to poor detection accuracy,a metric learning-based clustering algorithm is proposed.The algorithm generates different metric calculations according to the code samples,which has higher accuracy and generalization compared with the common clustering algorithm.To address the problem that the existing studied neural networks have low modeling ability for complex codes,which leads to poor stability,a scheme using Bi-Phased LSTM as a feature extractor is proposed,which uses loophole samples as time series to initialize the time gate of Phased LSTM and controls the update of memory units through the time gate,which effectively alleviates the problem of gradient disappearance and improves the framework stability.(2)A source code vulnerability detection system was designed and implemented,supporting C++and Python languages.Through the web interface,students and researchers can directly submit and detect source code vulnerabilities.The system will print the vulnerability information and highlight the location of the vulnerability,and provide suggestions to circumvent the vulnerability.Users can also edit configuration files to try different combinations of hyper parameters and quickly find the ideal model for new datasets Through the application programming interface,developers can also add features to the system and extend the languages for detection.