Adversarial Attack on Deep Learning Model Detecting SQLi in a Constrained Environment
Aswin Harish Kuttaten, Kurunandan Jain, Prabhakar Krishnan · 2025
SQL Injection (SQLi) attacks remain a significant threat to web applications, necessitating advanced detection mechanisms to combat evolving attack strategies. This paper proposes a novel approach to SQLi detection by generating adversarially perturbed payloads through a conditional perturbation strategy. Unlike traditional models, where adversarial samples are generated using a neural network, this study leverages a conditional perturbation to create realistic and context-specific malicious SQLi payloads such that the perturbation technique are based on mending SQLi rules and not to be burdened with performance overhead. The perturbation process selectively encodes characters and operators, mimicking real-world attack vectors. Since the approach is being evaluated within a Virtual Machine (VM) environment is to demonstrate its efficiency and robustness under constrained resources. Our results show that the adversarial perturbation technique significantly reduces the detection model’s effectiveness to classify between genuine and perturbed samples, hence lowering performance metrics such as accuracy, precision, recall, and F1-Score for SQLi detection. The findings highlight the vulnerability of detection systems to sophisticated adversarial attacks, emphasizing the need for resilient security measures in real-world applications.