Feature-Enhanced Deep Learning Models for Cyber-Physical System Security

The growing reliance on cyber-physical systems (CPS) in critical infrastructure has highlighted the need for accurate and efficient intrusion detection methods. Traditional anomaly detection techniques struggle to handle the complexity and variability of CPS data, often resulting in high false-positive rates. This study investigates the use of genetic algorithms (GA) with dynamic thresholding as a feature selection method to improve the performance of deep learning models. Three publicly available CPS datasets were used to validate the approach, including the water distribution testbed (WDT), power system, and gas pipeline datasets. Deep learning models, including Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), and Autoencoders, were enhanced with techniques such as attention mechanisms, evolving filters, and GA-based weight pruning. Overall, the integration of GAdriven optimization significantly boosted model accuracy and reduced loss across the board, average between one to three percent. The Hybrid CNN model optimized with Genetic Algorithms achieved the highest accuracy on the gas pipeline dataset, attaining a performance of $94.61 \%$. Likewise, the evolving filter model yielded the best results on the WDT dataset, with an accuracy of $94.13 \%$. This work lays the foundation for broader use of GA not only in feature selection but also in hyperparameter tuning for adaptive, efficient, and robust security systems.