Interpretable Feature Engineering for Robust Mobile Application Traffic Classification

Интерпретируемое конструирование признаков для устойчивой классификации трафика мобильных приложений
Abolghasem Rezaei Khesal, Mehdi Teimouri
2026-01-01

MING-ITC-2024TLS/QUIC metadataXGBoostencrypted mobile traffic classificationinterpretable feature engineering
Encrypted mobile traffic challenges traditional classification methods, demanding interpretable features that generalize across environments. We introduce 263 hand-engineered features from TLS/QUIC metadata, TCP sequence dynamics, intra-flow burst structures, and UDP first-byte patterns to capture application semantics without payload inspection. Using XGBoost on the ITC-2022 (controlled, single-app) and new MING-ITC-2024 (mixedapp, cross-regional) datasets, our features achieve 96–97% intra-dataset F1scores, outperforming traditional flow statistics, AppScanner, and Flowprint. Inter-dataset tests reveal that while behavioral features excel in validation, a compact server-feature set (IP, port, subnet) shows superior robustness, emphasizing stable infrastructure signals for real-world deployment.
1
A compact server-feature set using IP, port, and subnet provides superior cross-environment robustness, highlighting stable infrastructure signals for deployment.
2
Inter-dataset evaluation shows behavioral features perform strongly in validation but generalize less robustly across environments.
3
Introduces 263 interpretable, hand-engineered features from TLS/QUIC metadata, TCP dynamics, burst structures, and UDP first-byte patterns without payload inspection.
4
XGBoost achieves 96–97% intra-dataset F1 scores on ITC-2022 and MING-ITC-2024, outperforming traditional flow statistics, AppScanner, and Flowprint.

Encrypted mobile application network traffic

Interpretable feature representations and their cross-environment classification robustness, including the comparative value of behavioral versus stable server-side signals

Publication Details
Publication Date
2026-01-01
Journal
Publisher
ISSN
Cited by
1
Access Type
Author Information
Authors
Abolghasem Rezaei Khesal
Mehdi Teimouri
Explore further
Open the scid.ai AI chat with a ready-made request: it will find papers on a similar topic and help build a literature review.
Find similar papers in the chat →
Make a presentation
100%