Class imbalance remains a major challenge in developing machine learning models for SQL injection attack detection because harmful queries are usually underrepresented in real-world datasets. This study comparatively evaluates the effectiveness of classical, hybrid, and generative data augmentation techniques for improving SQL injection attack detection performance on imbalanced data. Seven augmentation methods—SMOTE, ADASYN, Borderline-SMOTE, SMOTE-Tomek Links, SMOTE-ENN, Generative Adversarial Networks, and conditional GAN- were applied to the SQLiV3 dataset and evaluated using Decision Tree, Random Forest, Naive Bayes, Support Vector Machine, and DistilBERT classifiers. Model performance was assessed using Accuracy, Precision, Recall, F1-score, Matthews correlation coefficient, and ROC-AUC. Results demonstrate that data augmentation significantly improves SQL detection compared with the original imbalanced dataset, although the degree of improvement depends on both the augmentation technique and classifiers. Among the conventional oversampling methods, SMOTE outperformed ADASYN, with Random Forest achieving a ROC-AUC of 98.16%. Among the hybrid techniques, SMOTE-ENN produced the best performance for conventional machine learning models, with Random Forest attaining 99.87% accuracy and a ROC-AUC of 99.67%. Generative Adversarial Network-based augmentation also achieved highly competitive results, with Random Forest recording 99.49% accuracy and a 99.82% ROC-AUC. Across all experiments, DistilBERT consistently outperformed conventional classifiers, achieving up to 99.87% accuracy and a ROC-AUC of 100% on the SMOTE-balanced dataset. Friedman’s statistical test confirmed statistically significant differences among the augmentation techniques. These findings demonstrate that augmentation methods that improve class balance and preserve discriminative characteristics of SQL injection queries produce superior detection performance.