Binary Function Similarity Detection Based on Graph Neural Network with Self-Attention Mechanism

Binary function similarity searching is an important method for vulnerability search and software security assurance. In order to meet the requirements of software security detecting on mobile platforms, we improve the high-accuracy graph matching algorithm and propose a SAGEN (Self-Attention based Graph Embedding Network) model. SAGEN can reduce retrieval time and improve retrieval speed while ensuring high accuracy, which enables high-accuracy graph matching algorithms to be applied to practical scenarios. The method first obtains the control flow chart by the binary function file, and then generates the graph data by the control flow chart. After that, this paper proposes a self-attention based graph neural network to embed graphs into a vector space. Finally, the similarity score between graphs is obtained by calculating the distance between graph embedding vectors. The experiment results show that SAGEN can achieve better results than most of existing models in accuracy while ensuring meeting retrieval speed requirements.