Detecting network intrusions in cyber-physical systems using deep autoencoder-based dimensionality reduction approach anddeep neural networks

Abstract

Cyber-Physical Systems (CPSs) that integrate computational and physical processes are the foundation of reliability in prominent areas of critical infrastructure, including transportation, energy, and manufacturing. The expansion in connected CPSs has made them vulnerable to various and changing intrusions into their networks. This research proposes a hybrid deep learning architecture that integrates the utilisation of a denoising autoencoder as a feature dimensionality reduction component with a five-layer deep feedforward neural network as an effective intrusion classifier. The model is trained and tested on CICIDS2017 and UNSW-NB15 datasets with a rich collection of attack patterns such as DoS, DDoS, Shellcode, and Worm attacks. The denoising autoencoder effectively learns higher-level representations of network traffic data, whereas the deep feedforward network facilitates precise multi-class classification. Empirical results demonstrate that the model achieves 99.99% and 99.95% detection accuracies on CICIDS2017 and UNSW-NB15 datasets, respectively, at very low false positive rates. Comparative analysis with state-of-the-art techniques further confirms the superior performance and generalisability of the presented solution, highlighting its applicability to real-time CPS threat detection systems.