Comprehensive Machine Learning Pipeline for Encrypted Traffic Classification With Bayesian Optimization and Explainability

Le résumé fourni par la source

Encrypted traffic classification is critical for modern cybersecurity, yet traditional approaches are ineffective and existing learning-based methods often struggle with noisy data, limited interpretability, and inefficient optimization. To address these challenges, this paper presents a comprehensive machine learning pipeline optimization framework enhanced with sequential model-based Bayesian optimization for robust and explainable encrypted traffic classification. The approach integrates explicit outlier detection, heterogeneous classifier benchmarking, and Bayesian optimization using the Tree-structured Parzen Estimator (TPE). It is evaluated on the BCCC-DarkNet-2025 dataset, which includes VPN, Tor, I2P, Freenet, and ZeroNet traffic, and is compared with state-of-the-art methods. Experimental results under repeated stratified within-dataset evaluation on the BCCC-DarkNet-2025 dataset show that the optimized pipeline achieves 99.37% accuracy, 98.63% F1-score, 0.982 MCC, and 98.15% Macro-F1 across 10 independent trials, with very low variance (std < 0.1%), indicating highly stable performance under this evaluation protocol. A comprehensive SHAP-based explainability analysis highlights key discriminative features, improving explainability, while sub-millisecond inference latency (std < 0.2 ms/sample) supports real-time deployment. Overall, the proposed framework demonstrates strong, stable, and explainable performance under repeated stratified evaluation on the BCCC-DarkNet-2025 dataset. Further validation using independent datasets and group-disjoint evaluation protocols remains an important direction for future work.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.