Resumen
Amid the incessant evolution of the Internet, an array of cybersecurity threats has surged at an unprecedented rate. A notable antagonist within this plethora of attacks is the SQL injection assault, a prevalent form of Internet attack that poses a significant threat to web applications. These attacks are characterized by their extensive variety, rapid mutation, covert nature, and the substantial damage they can inflict. Existing SQL injection detection methods, such as static and dynamic detection and command randomization, are principally rule-based and suffer from low accuracy, high false positive (FP) rates, and false negative (FN) rates. Contemporary machine learning research on SQL injection attack (SQLIA) detection primarily focuses on feature extraction. The effectiveness of detection is heavily reliant on the precision of feature extraction, leading to a deficiency in tackling more intricate SQLIA. To address these challenges, we propose a novel SQLIA detection approach harnessing the power of an enhanced TextCNN and LSTM. This method begins by vectorizing the samples in the corpus and then leverages an improved TextCNN to extract local features. It then employs a Bidirectional LSTM (Bi-LSTM) network to decipher the sequence information inherent in the samples. Given LSTM?s modest effectiveness for relatively long sequences, we further integrate an attention mechanism, reducing the distance between any two words in the sequence to one, thereby enhancing the model?s effectiveness. Moreover, pre-trained word vector features acquired via BERT for transfer learning are incorporated into the feature section. Comparative experimental results affirm the superiority of our deep learning-based SQLIA detection approach, as it effectively elevates the SQLIA recognition rate while reducing both FP and FN rates.