We investigate the problem of classifying a line of program as containing a vulnerability or not using machine learning. Such a line-level classification task calls for a program representation which goes beyond reasoning from the tokens present in the line. We seek a distributed representation in a latent feature space which can capture the control and data dependencies of tokens appearing on a line of program, while also ensuring lines of similar meaning have similar features. We present a neural architecture, Vulcan, that successfully demonstrates both these requirements. It extracts contextual information about tokens in a line and inputs them as Abstract Syntax Tree (AST) paths to a bi-directional LSTM with an attention mechanism. It concurrently represents the meanings of tokens in a line by recursively embedding the lines where they are most recently defined. In our experiments, Vulcan compares favorably with a state-of-the-art classifier, which requires significant preprocessing of programs, suggesting the utility of using deep learning to model program dependence information.
Download