Abstract
Security Information and Event Management systems are still at the core of threat monitoring in enterprises, but their rule-based detection methodology is mostly static in nature and always in need of manual tuning. As the nature of cyber-attacks becomes increasingly sophisticated, the complexity of the operating environment also increases, leading to a degradation in the accuracy of the rule-based methodology, with false positives and false negatives rising significantly. Recent studies show that adaptive learning methodologies can improve the accuracy of anomaly-based detection systems in a dynamic operating environment. Reinforcement learning is a methodology in which a learning agent learns through its interactions with its operating environment and improves its decision-making capabilities through a series of iterations. This research proposes a reinforcement learning-based framework for the continuous improvement of cyber threat detection rules in Security Information and Event Management systems. A reinforcement learning agent learns from the outcomes of the alerts generated in the system, the feedback from the Security Operations Center, and the threat intelligence available in the system to improve the thresholds and correlation values in real time for the rule-based methodology. This research uses benchmark intrusion datasets to evaluate the proposed methodology and compares its performance with static rulebased systems to show the improvement in accuracy and a reduction in false positives generated in the system.