What is it about?
The surge in mobile network usage has contributed to the adoption of Deep Learning (DL) techniques for Traffic Classification (TC) to ensure efficient network management. However, DL-based classifiers still face challenges due to the frequent release of new apps (making them outdated) and the lack of interpretability (limiting their adoption). In this regard, Class Incremental Learning and eXplainable Artificial Intelligence have emerged as fundamental methodological tools. This work aims at reducing the gap between the DL models’ performance and their interpretability in the TC domain. In this study, we examine from different perspectives the differences between classifiers when trained from scratch and incrementally. Using Deep SHAP, we derive global explanations to emphasize disparities in input importance. We comprehensively analyze base classifiers’ behavior to understand the incremental process’s starting point and examine updated models to uncover architectures’ features resulting from the incremental process. The analysis is based on MIRAGE19, an open dataset focused on mobile app traffic.
Featured Image
Photo by Igor Omilaev on Unsplash
Why is it important?
The primary focus of this paper is to leverage xAI analyses to uncover the inner mechanisms of incremental mobile traffic classifiers, as done in recent years for static viz. non-incremental classifiers, and contribute to the development of mobile networks that are more robust, reliable, and adaptable to the dynamic Accordingly, in this work (i) we investigate the differences (in predictions and input importance) between CIL-based classifiers and a model trained-from-scratch (abbr. scratch); (ii) we conduct a comprehensive analysis of the base classifiers' output to gain a thorough understanding of the initial conditions for the incremental process; (iii) we extensively examine various aspects of CIL-based traffic classifiers to elucidate the disparities between the updated classifiers and the scratch. The empirical studies are conducted on the MIRAGE dataset, which includes traffic generated by mobile applications, analyzing state-of-the-art CIL algorithms.
Perspectives
In this work, we investigated various perspectives on incremental traffic classifiers to gain a comprehensive understanding of their inner workings. By analyzing Deep SHAP explanations, we understood that fields and packets importance changes due to incremental training, (e.g., WIN becomes more important for YouTube). Unveiling the inner mechanism of updated models, we discerned that performance is influenced by similarities between some apps (e.g., Facebook-Messenger and YouTube-PlayStore), the bias toward the new apps stems from the stronger alignment between weights and feature vectors, and the correction layer strongly impacts on performance (e.g., Waze gains >30% F1-score while Facebook loses 28% F1-score). In future work, we plan to (i) leverage xAI insights to inform the design of incremental traffic classifiers and mitigate forgetting and intransigence, (ii) investigate and enhance the reliability of incremental traffic classifiers (via calibration analysis) and (iii) generalize the XAI analysis to other incremental techniques (e.g. model-growth).
Francesco Cerasuolo
Universita degli Studi di Napoli Federico II
Read the Original
This page is a summary of: Explainable Mobile Traffic Classification: the Case of Incremental Learning, December 2023, ACM (Association for Computing Machinery),
DOI: 10.1145/3630050.3630178.
You can read the full text:
Contributors
The following have contributed to this page
