Enhanced and Explainable Deep Learning-Based Intrusion Detection in IoT Networks

The proliferation of IoT networks has significantly increased the potential for cyber attacks. Deep learning models have shown effectiveness in detecting complex attacks; however, they face challenges related to imbalanced datasets and a lack of interpretability. In this work, we propose an enhanced and interpretable deep learning approach that addresses the common challenges of data imbalance and interoperability. To tackle the data imbalance issue, we employ CTGAN, a technique that expands the dataset by generating synthetic samples for the minority class traffic. Additionally, we utilize Boruta Shap for feature extraction, resulting in a reduced number of features and enhancing the efficiency of detection. Moreover, we incorporate SHAP for model explainability. We validate the results obtained from SHAP by conducting a thorough analysis of each attack type in both the NSL-KDD and UNSW-NB15 datasets. Furthermore, we conduct a comparative evaluation of our framework against a previous approach, demonstrating that our proposed framework outperforms the previous one in accurately detecting attacks for majority of class types.