Cyber-Physical Attack Detection in Water Distribution Systems Using a Hybrid Ensemble of XGBoost, Isolation Forest, and LSTM Autoencoder on the BATADAL Dataset

Abstract

Water distribution systems (WDSs) are critical public infrastructures increasingly controlled through cyber-physical layers, making them attractive targets for malicious intrusions. Real-time detection is difficult: confirmed attack data is scarce, sensor readings co-vary across dozens of channels, and well-crafted intrusions are deliberately kept within plausible operating ranges. This paper proposes a three-model hybrid ensemble that combines a supervised XGBoost classifier, an unsupervised Isolation Forest with principal component analysis (PCA) dimensionality reduction, and an unsupervised LSTM Autoencoder trained on 24-hour sliding windows of sensor sequences. All models are evaluated on the BATtle of the Attack Detection Algorithms (BATADAL) training dataset, which contains 4,177 hourly observations from a simulated SCADA-monitored C-Town network spanning July–December 2016, with 219 confirmed attack hours (5.2%) across five distinct attack campaigns. The weighted ensemble (XGBoost 40%, LSTM Autoencoder 50%, Isolation Forest 10%) achieves an F1-score of 0.9336, ROC-AUC of 0.9967, and an average precision of 0.9390, detecting 211 of 219 attack hours with only 22 false positives and a false positive rate of 0.56%. Time-to-detection analysis across all five attack windows confirms that the ensemble provides rapid and reliable detection, identifying four of five attack campaigns within one hour of onset. Results demonstrate that combining complementary detection paradigms substantially outperforms any single-model approach on this benchmark.