Comparative analysis of approaches to source code vulnerability detection based on deep learning methods

The object of research this work is the methods deep learning for source code vulnerability detection. One most problematic areas use only one approach in analysis process: based on AST (abstract syntax tree) or program dependence graph (PDG). In paper, a comparative two approaches detection was conducted: and PDG. various topologies neural networks were analyzed. They are used As result comparison, advantages disadvantages each determined, results summarized corresponding comparison tables. analysis, it determined that BLSTM (Bidirectional Long Short Term Memory) BGRU Gated Linear Unit) gives best terms problems showed, effective systems method uses an intermediate representation code, which allows getting language-independent tool. Also, work, our own algorithm system proposed, able to perform following operations: predict vulnerability, classify generate patch found vulnerability. A detailed proposed system’s unresolved issues provided, planned investigate future researches. could help speed up software development process as well reduce number vulnerabilities. Software developers, specialists field cybersecurity, can be stakeholders system.