Abstract
Anomaly Detection Systems (ADS) are essential in Industrial and Internet of Things (IIoT) environments by identifying equipment failures, environmental anomalies, operational irregularities, and cyberattacks. However, the increasing reliance on Machine Learning and Deep Learning (DL) exposes ADS to adversarial attacks, particularly transferable evasion attacks, where Adversarial Examples (AE) crafted for one model can deceive others. Despite their importance, limited research has examined the transferability of adversarial attacks in industrial and IoT contexts or the effectiveness of defense strategies against them. This work systematically evaluates the transferability of adversarial evasion attacks across six ADS models, including both tree-based and neural network architectures, trained on industrial and IIoT scenarios datasets. We also analyze multiple adversarial detection methods, measuring not only their performance, but also their computational efficiency in terms of execution time, processor utilization, and energy consumption. Our results show that most ADS are vulnerable to transferable evasion attacks and that existing detection methods fail in model- and attack-agnostic settings. We further demonstrate that incorporating adversarial learning with a small set of low-perturbation examples significantly improves detection while maintaining low computational overhead, enabling practical and efficient real-time deployment.