Research On Overflow Vulnerability Prediction Method Based On Multimodal Software Features

In recent years,with the rapid development of the software industry,the prediction and protection of software vulnerabilities have increasingly attracted more attention,making it a key research field in computer science.However,as the scale of software systems expands and architectures become increasingly complex,the difficulty of predicting software vulnerabilities is also increasing,especially for specific types of overflow vulnerabilities.Therefore,building accurate,effective,and practical prediction methods for overflow vulnerabilities has become an urgent need in the software industry.Considering the targeted nature of software overflow vulnerability features,the concealment of vulnerability information,and the high complexity of software features,this deeply analyzes the behavior patterns of overflow vulnerabilities in software source code and executable programs,constructs a multimodal software vulnerability feature composed of software graph structures,program slices,and disassembly semantics to fully extract vulnerabilities and preserve overflow vulnerability information.On this basis,effective prediction of various types of overflow vulnerabilities has been achieved.Firstly,to address the issue of predicting multi type overflow vulnerabilities using multiple modal features,a framework based on contextual semantic information for predicting multiple types of overflow vulnerabilities(MCSInf)is constructed.Based on the three different software features of software graph structure,program slices,and disassembly semantics,the framework constructs multimodal software features and utilizes the CBOW model to develop a feature transformation method.It also analyzes the direct causes of various types of overflow vulnerabilities in software and proposes an overflow vulnerability labeling method.The contextual semantic information network is selected as the prediction model for MCSInf,to achieve multi type overflow vulnerabilities targeting multiple modal feature prediction software.Secondly,based on the MCSInf framework,a prediction method based on updating source code graph structure is proposed to solve the problem of not being able to obtain overflow vulnerability-related features based on the graph structure of source code.Taking functions as the grains and based on code attribute graphs,a set of overflow vulnerability features called GSVFset is constructed,and a mechanism for obtaining overflow vulnerability features is formulated.The overflow vulnerability features are transformed into feature matrices based on the feature transformation method of MCSInf.An attention pooling layer with self-attention mechanisms is introduced,and a graph neural network with self-attention mechanisms is developed to update the overflow vulnerability feature based on the updated GSVFset feature matrices.The updated feature matrices are then labeled using the overflow vulnerability labeling method of MCSInf and input into the prediction model of MCSInf to predict multiple types of software overflow vulnerabilities based on the source code’s runtime behavior.Thirdly,based on the MCSInf framework,a prediction method based on feature combination slices is proposed to address the problem of not being able to fully obtain overflow vulnerability features of function source code.The method analyzes eight types of code behavior patterns that cause software to have overflow vulnerabilities,defines multiple types of overflow vulnerability features,sensitive functions,and IFS internal overflow vulnerability features,formulates feature extraction strategies and algorithms for extracting IFS internal overflow vulnerabilities,proposes EFS external overflow vulnerabilities,constructs overflow vulnerability combination slices,and proposes a transformation method for overflow vulnerability combination slices.Based on this,an embedding layer is constructed using the feature transformation method of MCSInf,and the labeled embedding layer is input into the prediction model of MCSInf using the overflow vulnerability labeling method of MCSInf to predict multiple types of software overflow vulnerabilities based on program slices.Then,based on the MCSInf framework,a prediction method based on disassembly semantic information is proposed to address the issue of insufficient acquisition of overflow vulnerability features in executable programs.The method analyzes the reverse-assembly semantic behavior patterns that cause various types of overflow vulnerabilities in software.A set of executable software overflow vulnerability features called ESVF is defined.For different features of ESVF,corresponding feature transformation methods and symbolic replacement custom function name methods are formulated.After feature transformation,they are concatenated to construct an embedding layer labeled based on the MCSInf’s overflow vulnerability labeling method.A prediction model based on multilayer perceptrons is constructed to predict various types of overflow vulnerabilities based on reverse-assembly semantic features.Finally,verify the performance of the constructed MCSInf framework and the proposed overflow vulnerability prediction method,four evaluation metrics are used:accuracy,precision,recall,and F1 value.The performance of the proposed method is analyzed on multiple types of software overflow vulnerability datasets,and the applicability,effectiveness,and accuracy of the constructed MCSInf framework and the proposed method are verified on operating systems,language libraries,and software.