Knowledge Distillation for SQL Injection Detection in IoT Edge Dashboards Using BiLSTM

Several studies have employed bidirectional long short-term memory (BiLSTM) for SQL injection detection, achieving satisfactory performance. However, the SQL injection solution on the IoT edge dashboard has not been optimized to consider the efficiency and real-time nature of these properties, let alone utilizing BiLSTM. This study aims to evaluate the performance of knowledge distillation, enabling BiLSTM to become more efficient and real-time in SQL injection detection on the IoT edge dashboard. We utilize the SQL injection detection dataset from Kaggle, which is reproducible for research purposes. We apply BiLSTM as a teacher model and then perform knowledge distillation to make the model more compact while monitoring the compression ratio (CR), accuracy loss, and prediction time. Finally, we utilize Flask to design a web-based user interface for detecting SQL injections. The test results show that knowledge distillation can provide optimal performance on BiLSTM, with accuracy, precision, recall, and F1-score values of 0.985, 0.985, 0.983, and 0.984, respectively, compared to the teacher and student models without a teacher. The efficiency of knowledge distillation is evident from the significant reduction in the size of the BiLSTM model, from 20,660.02 to 7,726, with a compression ratio (CR) of $2.67 \times$. Finally, applying knowledge distillation to SQL injection detection for IoT edge dashboards significantly reduces the prediction time from an average of 0.68 seconds to 0.17 seconds.