A Binary Vulnerability Similarity Detection Model Based on Deep Graph Matching

Abstract

To enhance network security, this study employs a deep graph matching model for vulnerability similarity detection. The model utilizes a Word Embedding layer to vectorize data words, an Image Embedding layer to vectorize data graphs, and an LSTM layer to extract the associations between word and graph vectors. A Dropout layer is applied to randomly deactivate neurons in the LSTM layer, while a Softmax layer maps the LSTM analysis results. Finally, a fully connected layer outputs the detection results with a dimension of 1. Experimental results demonstrate that the AUC of the deep graph matching vulnerability similarity detection model is 0.9721, indicating good stability. The similarity scores for vulnerabilities such as memory leaks, buffer overflows, and targeted attacks are close to 1, showing significant similarity. In contrast, the similarity scores for vulnerabilities like out-of-bounds memory access and logical design flaws are less than 0.4, indicating good similarity detection performance. The model’s evaluation metrics are all above 97%, with high detection accuracy, which is beneficial for improving network security.