A Deployment-Oriented Evaluation Framework for Robust SQL Injection Detection

<ns3:p>Background SQL injection attacks (SQLIA) remain among the most critical threats to web application security due to their ability to compromise database confidentiality, integrity, and availability. Although machine learning-based detection approaches have demonstrated high classification accuracy, many existing studies primarily focus on nominal performance metrics without adequately addressing robustness, statistical reliability, calibration behavior, or operational deployment constraints. Methods This study proposes a deployment-oriented evaluation framework for SQL injection detection using six supervised machine learning algorithms: Multinomial Naïve Bayes, Logistic Regression, LinearSVC, SGDClassifier, LightGBM, and XGBoost. All models were trained using TF-IDF weighted character-level n-gram representations. The framework integrates conventional classification metrics with bootstrap confidence intervals, McNemar statistical testing, adversarial robustness evaluation, calibration analysis, threshold sensitivity analysis, and operational performance profiling including latency and throughput assessment. Results Experimental results demonstrated that all evaluated models achieved near-ceiling classification performance under clean testing conditions, with F1-scores approaching 0.99. However, robustness, calibration reliability, and operational efficiency analyses revealed meaningful differences among classifiers. Character-level TF-IDF representations provided strong resilience against multiple SQL obfuscation strategies. Linear models demonstrated favourable computational efficiency and robustness stability, whereas ensemble methods showed stronger probability calibration behavior. Conclusions The findings indicate that SQL injection detection models should not be selected solely based on nominal accuracy improvements. A multi-dimensional deployment-oriented evaluation framework that incorporates robustness, statistical validation, calibration reliability, and operational feasibility provides a more reliable foundation for real-world adversarial deployment scenarios.</ns3:p>