Lightweight DL for Early SQLi Detection in Web Apps Security

Abstract

SQL injections are still one of the most significant threats to online applications.   It is hard to discover SQL injection attacks since SQL queries are always changing.   As these attacks become more complicated and happen more often, current methods of finding them that rely on keyword denial lists are becoming less successful.  To address this problem, we suggest a new model that turns text-based features into images.  This study introduces a Lightweight Deep Learning system for SQL Injection Attack Detection, designated as LWD-SQLIAS.  The system has three main steps: first, it uses NLP approaches to preprocess the dataset. Then, it uses a hybrid feature extractor that combines TF-IDF with the neural layers of VGG16 filters to extract features and create image features that are 244×224 pixels in size.   In the end, the proposed Lightweight MobileNetV2 (LW-MNV2) model, which uses the MobileNetV2 architecture and fine-tuning method, is used to optimize these feature images. The proposed LWD-SQLIAS was implemented on dataset 30919 of SQL injection samples obtained from Kaggle. The results show the LW-MNV2 performs better than the original MobileNetV2, achieving 99.85% accuracy, 99.74% precision, 99.86% recall, and 99.80% score