mVulSniffer: a multi-type source code vulnerability sniffer method

Abstract

Given the problem that the code slice used by existing deep learning-based vulnerability sniffer methods could not comprehensively encompass the subtle characteristics between vulnerability classes, and a single deep learning sniffer model had insufficient ability to learn long context-dependent information between cross-file and cross-function code statements, a multi-type source code vulnerability sniffer method was proposed.Firstly, fine-grained two-level slices containing the types of vulnerabilities were extracted based on the control dependency and data dependency information in program dependency graph.Secondly, the two-level slices were transformed into initial feature vector.Finally, a fusion model of deep learning vulnerability sniffer suitable for two-level slices was constructed to achieve accurate vulnerability detection of multi-type source code.The experimental results on multiple synthetic datasets and two real datasets show that the proposed method outperforms the existing multi-type source code vulnerability sniffer methods.

Keywords

• multi-type vulnerabilities sniffer;deep learning;attention mechanism;data dependency;control dependency