Integrating eXplainable Artificial Intelligence (XAI) and Deep Learning in Real-Time Network Intrusion Detection Systems: A Multi-Context Feature Shift Evaluation

ABSTRACT:

The exponential growth of network data and sophisticated cyber threats requires Intrusion Detection Systems (IDS) to possess both high accuracy and clear interpretability. This research proposes a comprehensive IDS framework that combines a hybrid Deep Learning architecture (CNN-LSTM) with the Synthetic Minority Over-sampling Technique (SMOTE) to efficiently identify anomalous behaviors. The focal point of this paper is the integration of the SHapley Additive exPlanations (SHAP) tool and adversarial evaluation approaches to break the "black-box" barrier of Deep Learning, providing transparent interpretations for every system decision. The system is empirically evaluated on modern benchmark datasets, demonstrating high performance in maintaining accuracy, enhancing human trust management, and tracking multi-context feature shifts.

Keywords: Intrusion Detection Systems (IDS), Explainable Artificial Intelligence (XAI), Deep Learning, SHAP, SMOTE, Concept Drift, Trust Management