An explainable three-way neural network approach for intrusion detection in IoT ecosystem

Deep neural networks (DNNs) have demonstrated remarkable potential in intrusion detection; however, the inherent uncertainty in the data often results in false alerts due to the deterministic nature of their probabilistic classification mechanism. In IoT environments, this can affect system reliability. Moreover, most of the DNN-based methods do not provide a transparent and interpretable output of the model. A secure, reliable, and accurate framework is urgently required to protect IoT systems. To fill this gap, this article introduces a novel, explainable three-way neural network approach called Ex3WNN. We introduce a 3WC mechanism in the final layer of the DNN using Game-theoretic Rough Sets (GTRS). This allows the model to handle uncertain cases more intelligently by distinguishing between confident decisions and those requiring further interpretation. By categorizing predictions into attack, suspicious (uncertain), and normal classes, 3WC helps manage uncertainty, ensuring that ambiguous cases are flagged for further analysis rather than misclassified. The GTRS framework is employed to determine optimal decision thresholds, which are derived while achieving the trade-off between generality and accuracy. This approach enhances both detection accuracy and reliability. Incorporating the suspicious region can considerably reduce false alerts and significantly enhance the reliability, security, and confidence of the intrusion detection system (IDS). Furthermore, we utilize the eXplainable AI (XAI) techniques to provide an interpretable and transparent model’s output. The experimental results from four relevant and comprehensive datasets show that the proposed method outperformed existing baselines.