ENCRYPTED TRAFFIC CLASSIFICATION IN 5G NETWORKS BASED ON EXPLAINABLE AI

This thesis investigates Machine Learning and Deep Learning techniques for encrypted traffic classification in 5G&Beyond networks, with a focus on achieving both high accuracy and model interpretability. A Random Forest and a Convolutional Neural Network are evaluated across multiple temporal observation windows. The experiments show that window duration significantly influences model performance, with distinct behaviors emerging between the two architectures: the Random Forest remains stable and easily interpretable, while the CNN captures finer temporal patterns but exhibits greater variability. SHAP-based explainability provides meaningful insights into the models’ decision processes, demonstrating that effective traffic classification can be achieved without relying on fully black-box systems. The study contributes to the development of transparent and trustworthy AI-driven network monitoring solutions.