A Machine Learning Framework for Dynamic Risk Assessment and Anomaly Detection in Database Security

As databases serve as critical infrastructure in modern information systems, ensuring their security has become increasingly important. Traditional risk assessment and anomaly detection methods often struggle with limited adaptability to dynamic threats, high false positive rates, and suboptimal real-time performance. This study proposes a novel machine learning framework for dynamic risk assessment and anomaly detection in database security. The framework integrates supervised and unsupervised learning algorithms to create an adaptive evaluation model capable of identifying abnormal behaviors and emerging threats with greater precision and efficiency—distinguishing itself from existing approaches. Compared with the RADS real-time anomaly detection model (Reference 5), which focuses on real-time monitoring of software-defined network traffic and relies primarily on unsupervised learning for anomaly identification, our framework adopts a supervised-unsupervised learning fusion logic: supervised models (e.g., optimized SVM, Random Forest) are used to learn known attack patterns from labeled data, while unsupervised models (e.g., enhanced K-Means, DBSCAN) capture unknown anomalies from unlabeled data. The two types of models are dynamically weighted based on data uncertainty (e.g., higher weights for supervised models when labeled data confidence is high, and vice versa), addressing RADS’ limitation of over-reliance on unsupervised learning and poor generalization to known threats. Experimental results demonstrate that the proposed model significantly outperforms traditional methods in terms of accuracy, recall, and computational efficiency. This research not only contributes a scalable and intelligent approach to database security but also offers practical applications in high-risk domains such as finance and cybersecurity.