Optimized Intrusion Detection Approach for Cyber‐Physical System Using Meta‐Learning With Stacked Generalization: An Ensemble Learning Inspired Approach

ABSTRACT Cyber‐physical systems (CPSs) are crucial in providing vital infrastructure like smart grids, smart cities, smart automobiles, smart healthcare systems, and so forth, for many nations. CPSs are vulnerable to various attacks due to their large attack surface. An attack on these systems may lead to the disruption of critical services. To protect these systems, an optimized intrusion detection approach is needed. Although many intrusion detection approaches for CPSs exist, they have limitations like poor accuracy, high detection time, high space and time complexities, and high false alarm rates, etc. An optimized stack generalized meta‐learner‐based intrusion detection approach for CPSs has been proposed in this paper. The proposed approach utilizes numerous core models and a meta‐learner to classify network traffic in CPSs. The base models are trained on the learning data, and their outcomes are used as input features for the meta‐learner, which then makes the final prediction. Four classifiers are being used as base models, namely random forest (RF), gradient boosting (GB), multiple layer perceptron (MLP), k ‐nearest neighbors (KNNs), and extreme gradient boosting (XGB) classifier is used as a meta‐learner. The final predictions are generated using a stacking ensemble approach. Auto encoders are used for feature extraction, thereby utilizing a unique objective function designed for recursive attribute elimination. The presented approach selects only 10 out of the 46 features, which helps in reducing the space and time complexities. While implementing the proposed approach on the CIC‐IoT‐2023 dataset, the following results are obtained: multi‐classification accuracy (98.94%), precision (0.99), recall (0.99), F 1 score (0.99), average false positive rate (0.0003), detection time (0.12 s). When implemented using the NSL‐KDD dataset, the following results are obtained: multi‐classification accuracy (99%), precision (0.99), recall (0.99), F 1 score (0.99), average false positive rate (0.0012). While using the UNSW‐NB15 dataset, the following results are obtained: multi‐classification accuracy (99.56%), precision (0.99), recall (0.99), F 1 score (0.99), and average false positive rate (0.0002). The proposed approach performs better in contrast to the other cutting‐edge approaches. Also, the proposed approach introduces a novel and effective strategy for intrusion detection in CPSs.