DBPrompt: A Database Anomaly Operation Detection and Analysis via Prompt Learning

Database systems are widely used in various domains to store a significant amount of critical data. However, there has been a growing incidence of anomalous database access behaviors, causing significant impacts on database security. Despite the array of contemporary database protection methods, the existing methods have the following limitations: (1) Limited Interpretability. (2) Traditional methods of anomaly root cause analysis lack automation and rely on manual identification, leading to various misjudgments. (3) The scale of the database operation logs is enormous, requiring substantial resources for training. We introduce DBPrompt, a new semi-supervised and interpretable method for detecting and analyzing database anomalies. It reduces resource usage and prevents data forgetfulness in large language models through clustering and iterative summaries. DBPrompt uses prompt learning with specialized strategies for detecting database anomalies. Our tests show that DBPrompt outperforms current methods without needing fine-tuning and effectively classifies and analyzes identified anomalies, yielding positive results.