<![CDATA[This study uses the BERT architectural technique to identify schema exfiltration in a neural network-based Text-to-SQL system. The growing usage of Large Language Models (LLM) in Text-to-SQL systems, which may provide a danger of database schema leaking through user prompts, provides the context for this study. This research challenge is how to use a deep learning model to reliably and adaptively identify prompt modifications that could carry out exfiltration techniques. The study employed a deep learning strategy with a feedforward neural network as the classifier and the BERT architecture as the primary encoder. There were 20 classes in all, consisting of 19 exfiltration scheme categories and 1 benign class. The dataset was created using a variety of sources, including WikiSQL, DatabaseAnswers, and educational datasets. It was then subjected to tokenisation, labelling, and normalization processes. The model obtains an accuracy of 0.9462, precision of 0.8425, recall of 0.7483, F1-score of 0.7926, and precise match accuracy of 0.7596, according to the data. Additionally, the study demonstrated that the model outperformed implicit suggestions like role switching and prompt injection in identifying explicit prompts. The study concludes that while there are still issues with enhancing detection capabilities for intricate manipulating patterns, the BERT-based approach can provide good performance in identifying schema exfiltration in Text-to-SQL systems.]]>
