AN ENSEMBLE OF TOKENIZATION, AUTOENCODER AND TEMPORAL CONVOLUTIONAL NETWORK FOR SQL INJECTION DETECTION.

SQL Injection (SQLi) attacks remain a critical cybersecurity threat, with traditional detection methods struggling against novel attack patterns. This study proposes a hybrid deep learning architecture combining SQL-aware tokenization, a 1D Convolutional Autoencoder (1D-CAE), and a Temporal Convolutional Network (TCN) for robust SQLi detection. The framework leverages the autoencoder for unsupervised anomaly detection and the TCN for temporal sequence modeling. Evaluated on 30,918 SQL queries, the proposed ensemble achieved 98.53% accuracy, 92.5% precision, 95.5% recall, and 94.0% F1-score, significantly outperforming Random Forest (87.0% F1-score) and individual model components. The TCN's parallel processing capability enabled 12.3 milliseconds average inference time per query, supporting real-time deployment. The fusion of anomaly-based and sequence-based deep learning provides an efficient, scalable defense against both known and zero-day SQLi attacks.