Towards lightweight model using non-local-based graph convolution neural network for SQL injection detection

SQL injection poses serious threats to web applications and databases by enabling unauthorized access and data leakage. To address this issue, we propose a unique graph network, an innovative topology not explored previously for SQL injection detection. SQL statements are nodes, and their connections form edges in the graph. We introduce three graph CNN models, including a graph classification model with a two-layer Graph Convolutional Network (GCN), a graph classification model leveraging a non-local graph convolution derived from a 1x1 convolution, supplanting the original 1x1 convolution, and a modified non-local-block module by substituting the 1x1 convolution layers in the non-local architecture with GCN. The proposed models exhibit accuracy above 99% and inference times under 1 ms on two datasets. In comparison with traditional 22 models, our models using GCN demonstrate superior computation efficiency, parameter reduction, accuracy enhancement, and the advantage of handling input sequences of any length, underlining their potential in real-world cybersecurity systems, especially in effective SQL injection detection and mitigation strategies.