A simple function embedding approach for binary similarity detection

Binary function similarity detection has been an important problem in binary analysis. Recently, several deep learning based techniques show promising results and achieve state of the art performance. In despite of effectiveness, some of them adopt complex network structures which result in slow convergence speed. To reduce the structure complexity, this paper proposes a simple neural network structure based on Bi-RNN(Bidirectional Recurrent Neural Network). To evaluate the effectiveness of proposed method, this paper conducts experiments on both single-architecture and cross-architecture function similarity detection over OpenSSL library. Experimental results show that our approach achieves higher AUC metric and faster convergence speed than GNN(Graph Neural Network)based techniques. To explore the effectiveness of GNN, another experiment is conducted to explore the impact of the number of GNN layers on performance and the result shows that when combined semantic information, existing GNN-based methods may reduce performance.