A Robustness and Computational Efficiency Evaluation Pipeline For Machine Learning-Based Traffic Classiffication In 5g Network Slicing

Network slicing represents a core technology to improve heterogeneous Quality of Service (QoS) in 5G and beyond. However the effective network slicing implementation is challenging due to the di verse QoS requirements and dynamic conditions in 5G networks. Machine learning techniques have emerged as a viable solution to mitigate network slicing complexity in 5G. Moreover, the use of machine learning techniques introduces potential vulnerabilities that can lead to severe security risks. In this paper, we propose a structured ve-stage pipeline to analyze the trade-o s between predictive performance, computational efficiency, and adversarial robustness in MLP-based 5G traffic classification. Feature selection using Mutual Information and PCA reduced the original feature set by more than 50% while preserving accuracy. Hyperparameter optimization further decreased CPU usage and training time with minimal performance loss. However, adversarial evaluation under FGSM and PGD attacks revealed a signi cant degradation in performance as the intensity of the perturbation increased. The adversarial training strategies improved robustness at di erent levels of perturbation, albeit with increased computational cost.