STEWART: STacking Ensemble for White-Box AdversaRial Attacks Towards more resilient data-driven predictive maintenance
Onat Güngör, Tajana Rosing, Barış Akşanlı
Abstract
Industrial Internet of Things (I-IoT) is a network of devices that focus on monitoring industrial assets and continuously collecting data. This data can be utilized by Machine Learning (ML) methods to perform Predictive Maintenance (PDM) which identifies an optimal maintenance schedule for the industrial assets. The computational systems in the I-IoT are usually not designed with security in mind. Their limited computational power creates security vulnerabilities that attackers can exploit to prevent asset availability, sabotage communication, and corrupt system data. In this work, we first demonstrate that cyber-attacks can impact the performance of ML-based PDM methods significantly, leading up to 120 × prediction performance loss. Next, we develop a stacking ensemble learning-based framework that stays resilient against various white-box adversarial attacks. The results show that our framework performs well in the presence of cyber-attacks and has up to 60% higher resiliency compared to the most resilient individual ML method.