<![CDATA[Computer network security has become increasingly important with the growing complexity of cyberattacks. Deep learning-based Intrusion Detection Systems (IDS) represent a potential solution due to their capability to capture sequential patterns in network traffic. This study proposes a Double-Layer GRU-based IDS with Feature Fusion to enhance the representation of both numerical and categorical data in the NSL-KDD dataset. The training process employs systematic preprocessing techniques, including normalization and one-hot encoding. Experimental results demonstrate high accuracy and generalization with stable performance on both training and testing data, as well as competitive macro F1-scores for multi-class attack detection. Furthermore, interpretability aspects are explored through Explainable Artificial Intelligence (XAI) methods using SHAP and LIME. SHAP provides global insights into the contributions of important features, while LIME explains the influence of features at the local level for individual predictions. The integration of both methods not only enhances transparency and trust in the IDS but also offers deeper insights into dominant attributes in detecting attack patterns. Accordingly, this study contributes to the development of IDS that are accurate, interpretable, and applicable to modern network security.]]>
Copyrights © 2025
