Machine Learning Approach for Adaptive Data Protection in Document Database Systems

Abstract

With the growing dependence of cloud applications and enterprise organizations on NoSQL databases, it is necessary to ensure data protection while maintaining efficiency and performance. Traditional static encryption systems deliver strong regulatory protection but are unable to react to changing contexts in dynamic and zerotrust environments, which limits their ability to address abnormal behaviour, malicious insiders, and advanced attackers. This work proposes AdaptiCrypt-ML, a lightweight proxy based on machine learning, which aims to implement domain-level adaptive encryption in NoSQL database security systems. The framework utilizes the LightGBM model to classify 14 contextual, behavioural, and data-sensitivity features to determine immediately the most appropriate encryption level across four security categories. When data is entered, the encryption level is dynamically determined according to the risk level, whereas a risk-based decryption policy controls the extent to which data is revealed when retrieved. Empirical results, derived from a statistically validated synthetic dataset of 50,000 examples, demonstrate strong predictive performance, with an overall accuracy of 99.1%, an F1-macro score of 0.963, and a low generalization gap of 0.0018. The average inference time ranged between 0.5 and 0.8 milliseconds, and the total response time stabilized at 3.25 milliseconds (P95 = 4.10 milliseconds), with an average of 3,120 queries per second. A 5% noise robustness test validated 96% performance stabilization. These findings emphasize the possibility of integrating context-aware adaptive encryption into NoSQL frameworks without sacrificing real-time requirements.