Toward Efficient Network Traffic Classifications via Multimodal Learning

With revolutionized various real-time and cyber-physical applications, today’s networked systems face more threats of cyberattacks due to their expansive networking attack surface. One promising trend is to deploy deep learning network (DNN) models to classify and identify cyberattacks at the network layer to identify either malicious traffic or anomaly traffic as alerts. However, deploying DNN models to classify network traffic will compromise the efficiency requirements of networked systems since high-performance DNN models are usually complex at inference. With more attack surfaces in today’s networked system, more complex DNN models are required to fulfill the feature extraction. In this paper, we propose a multi-model learning approach for both accurate and efficient traffic classification to address the above challenge. Our key insight is that network traffic is multi-model data including structured data (e.g., protocol text) and non-structured data (e.g., packets). By designing different lightweight feature extractors, we extract features of multi-modal data via language models and linear models respectively, and fuse features from heterogeneous network traffic data to classify and identify typical cyberattacks. Experimenting with showcases of adopting classic learning methods indicates that our approach can achieve an accurate and efficient network traffic classification to ensure the security of networked systems.