Optimized Deep Learning Techniques for Intrusion Detection with Explainable AI-Based Ensemble Feature Selection

Using the idea of a honeynet, this study aims to build a realistic replica of an industrial setting in order to lure and capture would-be attackers. Hackers unintentionally reveal their strategies, methods, and processes by penetrating this honeypot’s structure, which helps with awareness of situations and detection of intrusions. To apply the XIDS to industry cyber-physical systems (ICPS), this study developed two new stages of approach. These approaches include the chi-square info gain, and gain ratio. The suggested approach improves the efficacy and precision of ICPS threat detection by picking discriminative characteristics. For detecting breaches using ICPS, the Bi-directional Long Short-Term Memory Model (BiLSTM) model has been fine-tuned and optimized. The models that were trained on a large dataset of ICPS traffic data, have a better capacity to distinguish between benign and illicit behaviour, which improves the effectiveness of detection and reduces false positives. Hyper-parameters adjustment in the Bi-LSTM model is carried out using the Bayesian optimization approach in this manner. It follows a comparison with the GRU and LSTM single classifiers. Performing better than competing models like GRU, LSTM, and BI-LSTM, the suggested BO-BI-LSTM approach attains a remarkable total accuracy of $99.42 \%$. The remarkable precision and resilience of detection of intrusions in manufacturing cyber-physical structures are amplified by the combined efforts of Explainable AI-based Ensemble Filter Feature Selection (XAI-EFFS) and Bayesian optimizing, which effectively fine-tune the model’s hyperparameters.