Abstract:Both traditional neural network models and graph neural network models have been validated as effective methods in the field of software vulnerability detection, but most of the current solutions are aimed at source code for vulnerability detection, and relatively little research has been conducted on applying neural network models directly to binary software for vulnerability detection, especially graph neural networks for binary software vulnerability detection is lacking. In order to fully investigate the effectiveness of neural network models in binary software vulnerability detection, this paper proposed a composite neural network based binary software vulnerability detection method. Firstly, the binary code was vectorised into a graph data structure that supports both traditional neural network models and graph neural network models for training. Then, a composite neural network model that combines traditional neural network models and graph neural network models was used to learn and validate the graph data structure. Finally, experiments and comparative analysis were conducted on a publicly available binary software vulnerability dataset. The experimental results show that the method could effectively improve vulnerability detection capabilities, with significant improvements in performance metrics such as accuracy and precision.
