A Comparative Study of GamiNet and Tree-Based Models for VPN and Non-VPN Traffic Classification

Accurate classification of VPN and non-VPN network traffic is critical for network management, security monitoring, and Quality of Service optimization. Thisstudy presents a comparative analysis of GAMI-Net, an interpretable deep learningframework, and state-of-the-art tree-based models for VPN versus non-VPN trafficclassification. Experiments were conducted using the ISCX VPN–nonVPN 2016dataset, comprising 123 PCAP files across 16 application categories, representingdiverse traffic behaviors such as streaming, messaging, and file transfer. Flow-levelfeatures were extracted using Tranalyzer2, followed by structured preprocessing,labeling, and correlation-based feature reduction, resulting in a compact set of 81informative features for model training. GAMI-Net leverages additive main effects and pairwise interactions to balance predictive accuracy with interpretability,while hyperparameters were optimized via Optuna using 3-fold stratified crossvalidation. Performance evaluation demonstrates that GAMI-Net achieves an accuracy of 97.59%, a precision of 90.16%, a recall of 98.27%, and an F1 score of93.71%, highlighting competitive performance compared to tree-based ensemblemethods such as Random Forest, XGBoost, and LightGBM. While slightly lowerin overall accuracy, GAMI-Net provides the added advantage of transparency, allowing insight into feature contributions and interaction effects. These findingssuggest that GAMI-Net is a viable approach for reliable and interpretable VPNtraffic classification, offering a practical trade-off between accuracy and explainability.