Adversarial Robustness of URLLC Traffic Classifiers using Network KPI Features

Ultra-Reliable Low-Latency Communication (URLLC) services rely on precise monitoring and timely prediction of network Key Performance Indicators (KPIs). In recent years, machine learning models have been widely adopted to interpret these KPIs, enabling automatic classification of network conditions and early detection of reliability degradation. However, an important concern is that such models can be sensitive to subtle, intentional perturbations in the input data. In this paper, we examine the adversarial robustness of a KPIdriven URLLC traffic classifier built using the Unicorn-Genesys dataset. The raw experiment logs are first converted into compact feature representations by aggregating KPIs through statistical measures such as mean, standard deviation, and maximum values. These features are then used to train a multilayer perceptron (MLP) model for reliability prediction. To assess robustness, we subject the classifier to adversarial perturbations generated using the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD). Our findings indicate that the classifier performs strongly under normal conditions, achieving close to 98% accuracy. However, even small adversarial perturbations lead to noticeable performance degradation, with iterative PGD attacks having the most severe impact. We further explore KPI-level sensitivity using a vulnerability map, which highlights the measurements most prone to manipulation. Adversarial training is also explored to enhance model robustness against attacks. Overall, the results point to potential security risks in data-driven network management and emphasize the need for more robust training and monitoring strategies.