Adversarial testing of machine learning models designed for SQL injection detection

Abstract

SQL injection remains one of the most prevalent web application vulnerabilities, consistently ranked among the most critical security threats according to the OWASP classification. In response to the increasing complexity of such attacks, contemporary research and engineering approaches have actively adopted machine learning (ML) and deep learning (DL) techniques for the automated detection of malicious SQL queries. Despite their high classification accuracy on static datasets, these models are susceptible to adversarial attacks—intentionally crafted inputs that preserve both the syntactic validity and semantic equivalence of the original queries but disrupt the model’s decision logic. This paper presents a method for adversarial testing of machine learning models designed for SQL injection detection. The proposed approach is based on the sequential application of mutation operators that maintain the syntactic correctness and semantic equivalence of the input queries. The algorithm is implemented as an iterative process aimed at reducing the model’s confidence in identifying a malicious query, while operating under black-box constraints without access to internal parameters or architecture. The method has been experimentally validated on three model architectures: MobileBERT, LSTM, and a hybrid CNN-LSTM. The evaluation results demonstrate a notable decrease in precision and recall for the injection class after applying adversarial mutations. The proposed framework serves as an effective tool for objective assessment of model robustness against polymorphic SQL injections and provides a foundation for developing more resilient security systems. Due to its universality and reproducibility, the method is also applicable for augmenting training datasets in adversarial learning settings. Furthermore, it holds potential for adaptation to other types of text-based attacks (e.g., XSS) and integration into industrial-grade web application security solutions.