Zero-Shot Traffic Identification with Attribute and Graph-Based Representations for Edge Computing.

Area of Science:

• Computer Science
• Network Security
• Machine Learning

Background:

• Fine-grained traffic identification is crucial but challenging due to mobile device proliferation and network growth.
• Existing machine learning and deep learning methods struggle with unseen traffic data due to reliance on training data distribution.

Purpose of the Study:

• To propose AG-ZSL, a zero-shot learning framework for general encrypted traffic classification.
• To address the limitations of current methods in handling unseen traffic samples.

Main Methods:

• AG-ZSL learns two mapping functions: traffic behavior embeddings from interaction graphs and attribute embeddings from descriptions.
• It minimizes the distance between these embeddings in a shared feature space.
• A two-stage classification method using gradient rejection and K-Nearest Neighbors is employed.

Main Results:

• AG-ZSL demonstrates exceptional performance in classifying both known and unknown traffic types.
• The framework shows high effectiveness on Internet of Things (IoT) datasets.

Conclusions:

• AG-ZSL offers a robust solution for general encrypted traffic classification, particularly for unseen samples.
• The framework has significant potential for enhancing secure and efficient network traffic management at the network edge.