Intrusion Detection Simplified: A Feature-free Approach to Traffic Classification Using Transformers

Kohei Miyamoto, Chansu Han, Tao Ban, Takeshi Takahashi, Jun'ichi Takeuchi

Abstract: Machine learning-based intrusion detection systems have emerged as powerful tools for defending against cyber attacks by automatically learning detection patterns from network traffic. Traditionally, these systems rely on manually designed flow-level features. However, this approach is limited by its dependence on domain expertise and the inability to quickly adapt to new attack vectors. As cyber threats evolve, there is a growing need for more flexible and scalable methods that can handle diverse and dynamic traffic patterns without extensive feature engineering. In this paper, we propose a transformer-based method for traffic classification for intrusion detection. Our model processes raw packet sequences, using their basic information including timestamp, while employing the time2vec technique to capture temporal characteristics in the data. This approach allows the model to learn representations of network traffic useful for intrusion detection without manually designed features. We evaluated the proposed method on a benchmark dataset and found it to be effective in detecting various kind of attacks in the dataset.