A Framework for Drift Detection and Adaptation in AI-driven Anomaly and Threat Detection Systems

A Framework for Drift Detection and Adaptation in AI-driven Anomaly and Threat Detection Systems

연구 분야: Safety

학회: International Journal of Information Security

The dynamic and evolving nature of cybersecurity threats presents significant challenges to anomaly and threat detection systems, particularly those that rely on Artificial Intelligence (AI) as their detection engine. A key limitation of current AI models is their inability to adapt to concept drift, feature drift, and adversarial attacks, which degrade performance over time. Although these phenomena arise from different underlying processes, they all share the effect of misaligning the operational data with the model’s training data. This study introduces the Hybrid Drift Detection and Adaptation Framework (HDDAF), which is a multi-layered AI system that is specifically designed to mitigate concept drift, feature drift, and adversarial attacks in cybersecurity. By framing all three challenges, HDDAF provides a unified approach that detects and responds to both natural evolution and malicious manipulation within a single adaptive pipeline. HDDAF integrates Hoeffding drift detection, feature selection, adversarial training, and incremental learning, allowing it to dynamically adapt through a Mixed-Drift Handling Module, which balances fine-tuning and full retraining. On the CIC-IDS2017 dataset, HDDAF achieves a macro F1 score above 99% and in tests on related datasets, it consistently adapts to data shifts with minimal retraining. An ablation study confirms that each module contributes to overall robustness, and real-time simulations demonstrate its ability to process high-velocity streams with stable latency and resource use. HDDAF’s hybrid design delivers both high accuracy and scalable performance for real-world cybersecurity applications.