Carbon-Aware Adversarial Detection in IIoT via Projection Transform

In the Industrial Internet of Things (IIoT) environment, the integration of the Internet of Things (IoT) and Artificial Intelligence (AI) facilitates various applications. Sensors deployed in critical areas continuously collect diverse data, including real-time air quality monitoring, to provide accurate environmental insights. However, these systems are susceptible to adversarial examples (AEs), including physical AEs. These attacks can compromise the accuracy of predictions and lead to misinterpretation of carbon emission levels. For instance, malicious factory owners could exploit AEs to evade carbon emission inspections. To address this challenge, this study introduces a method that utilizes Poisson distribution-based statistical modeling to detect AEs by analyzing distinct neuron activation patterns. Furthermore, to address the complexity and variability of air quality data, we introduce a vector projection technique to enhance the alignment of probability vectors with the model’s actual output feature space. Experimental results demonstrate the effectiveness of this method in detecting attacks generated by FGSM, PGD, DeepFool, and C&W, achieving F1-scores of 0.993, 0.992, 0.972, and 0.961, respectively. This enhances the reliability and energy efficiency of IoT-based air quality monitoring systems within carbon-intelligent IIoT framework.