Process-aware ML based intrusion detection for industrial control systems

Mugenzi Byiringiro, Raphaƫl
(2025)

Files

MugenziByiringiro_32421600_2025.pdf
  • UCLouvain restricted access
  • Adobe PDF
  • 3.59 MB

Details

Supervisors
Sadre, Ramin
Faculty
Ecole polytechnique de Louvain
Degree label
Master [120] : ingƩnieur civil en mathƩmatiques appliquƩes, Ơ finalitƩ spƩcialisƩe
Abstract
enIndustrial Control Systems (ICS) play a crucial role in critical infrastructure and must be protected from malicious intrusions that could disrupt real-world processes. This thesis explores a Machine Learning-based Intrusion Detection System (IDS) that relies on robust yet adaptable techniques, without requiring system-specific knowledge. Using the Secure Water Treatment (SWaT) dataset developed by iTrust at the Singapore University of Technology and Design, we propose a data-driven approach focused on signal classification and prediction. First, signals are preprocessed and analyzed to determine their cyclic or acyclic nature, leveraging spectral features and supervised learning. Then, predictive models such as ARIMA and LSTM autoencoders are trained on normal behavior to detect anomalies during attack scenarios. Evaluation shows that while ARIMA offers interpretability and speed, LSTM provides superior accuracy and reactivity. Our results highlight the potential of combining signal-type awareness with time-series prediction for effective intrusion detection in ICS environments.