Minimal Adversarial Attacks on Predictive Maintenance

In the era of Industry 4.0, predictive maintenance has become a key component of this new industrial paradigm. Predictive maintenance systems are generally based on deep learning models and aim to detect anomalies to anticipate future failures of industrial equipments. These systems have become a crucial security feature, monitoring sensor data in real time and detecting any abnormal behavior caused by human actions or malware. However, deep learning methods have been previously demonstrated being affected by intentional perturbations and vulnerable in adversarial settings. As a key component of industry security, predictive maintenance models are also impacted. In this context, adversarial attacks like FGSM and BIM have already been studied, highlighting the limitations of deep learning based predictive maintenance systems. This paper introduces two new attacking methods on predictive maintenance models. These attacks aim to minimize the transformations applied to adversarial instances focusing on the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$L_{2}$</tex> norm and the number of affected sensors. We benchmark our attacks on a total of six models trained on the NASA Turbofan Jet Engine dataset, demonstrating superior performances compared to existing FGSM and BIM attacks. Our results show that applying a modification with an average RMSE less than 0.05 allows us to generate adversarial examples with predicted Remaining Useful Life (RUL) close to 0 and high transferability properties. Additionally, we demonstrate that on the most advanced models, controlling only one or two sensors is sufficient to deceive predictive maintenance systems.