基于时空特征融合的SQL注入检测研究  

Research on SQL Injection Detection Based on Spatiotemporal Feature Fusion

作　　者：王清宇 王海瑞[1] 朱贵富 孟顺建 WANG Qing-yu;WANG Hai-rui;ZHU Gui-fu;MENG Shun-jian(Faculty of Information Engineering and Automation,Kunming University of Science and Technology)

机构地区：[1]昆明理工大学信息工程与自动化学院

出　　处：《化工自动化及仪表》2023年第2期207-215,共9页Control and Instruments in Chemical Industry

基　　金：国家自然科学基金项目(61863016,61263023)。

摘　　要：针对深度学习方法检测SQL注入时特征提取效果欠佳的问题,提出一种基于时空特征融合的检测模型SFFM。首先使用BERT预训练模型进行词嵌入,使用TextCNN提取SQL样本中不同粒度下的局部空间特征,同时使用BiGRU在保证训练效率的同时提取SQL样本的时序特征;再把提取到的特征送入Attention层进行全局语义信息提取;最后将提取到的特征进行融合,连接全连接层后送入softmax分类器进行分类检测。对比实验结果表明:SFFM模型获得了高达99.95%的准确率和99.90%的召回率,相较于CNN、LSTM和BERT-base模型,具有更好的检测效果。Considering poor feature extraction effect in detecting SQL injection through employing the deep learning method,a SFFM(spatiotemporal feature fusion model)-based detection model was proposed. In which, having BERT pre-training model used for word embedding and TextCNN employed to extract local spatial features of SQL samples at different granularity;meanwhile, having BiGRU adopted to extract temporal features of the SQL samples while ensuring a training efficiency;then, having the extracted features sent to the attention layer for global semantic information extraction;finally, having the extracted features fused and connected to the full connection layer and sent to the softmax classifier for classification detection. A comparative experiment shows that, the SFFM-based detection model can achieve an accuracy rate of 99.95% and a recall rate of 99.90%, and the SFFM-based detection model, as compared to CNN,LSTM and BERT-base models with single or simple structure, has better detection effect.

关 键 词：SQL注入检测 时空特征融合 SFFM模型 注意力机制 词嵌入 

分 类 号：TP393.08[自动化与计算机技术—计算机应用技术] TP181[自动化与计算机技术—计算机科学与技术]