Transformer-based tokenization for IoT traffic classification across diverse network environments.

Area of Science:

• Computer Science
• Network Security
• Machine Learning

Background:

• Internet of Things (IoT) traffic expansion necessitates accurate classification for network security and efficiency.
• Existing methods struggle with generalization, data dependency, and dynamic scenarios.
• Transformer models show promise but have limitations with irregular IoT traffic and single-task confinement.

Purpose of the Study:

• To introduce MIND-IoT, a scalable framework for generalized IoT traffic classification.
• To address limitations of current methods in handling diverse IoT environments and data.
• To develop a robust and adaptable solution for real-world IoT network challenges.

Main Methods:

• A hybrid architecture combining Transformer models and Convolutional Neural Networks (CNNs).
• IoT-Tokenize: A custom tokenization pipeline for preserving network flow semantics.
• Two-phase operation: Pre-training using Masked Language Modeling (MLM) and task-specific fine-tuning.

Main Results:

• Achieved up to 98.14% accuracy and 97.85% F1-score across diverse datasets.
• Demonstrated superior performance, robustness, and adaptability compared to traditional methods.
• Showcased ability to classify new datasets and adapt to emerging tasks with minimal fine-tuning.

Conclusions:

• MIND-IoT offers a highly effective and scalable solution for IoT traffic classification.
• The framework's hybrid architecture and custom tokenization enhance generalization and efficiency.
• MIND-IoT represents a significant advancement in securing and optimizing IoT networks.