Multimodal fusion based few-shot network intrusion detection system.

Area of Science:

• Cybersecurity
• Machine Learning
• Network Security

Background:

• Increasingly complex network environments and evolving attack vectors necessitate advanced intrusion detection systems.
• Traditional methods struggle with limited labeled data for new or rare network attacks, hindering few-shot learning effectiveness.
• Existing few-shot learning models often rely on single-modality data, failing to leverage complementary information across diverse data types.

Purpose of the Study:

• To develop a multimodal fusion based few-shot network intrusion detection method.
• To address the challenge of limited training data for detecting novel and rare network attacks.
• To improve network intrusion detection performance by integrating diverse data modalities.

Main Methods:

• Proposed a multimodal fusion approach combining traffic feature graphs and network feature sets.
• Developed a G-Model using Convolutional Neural Networks for traffic feature graphs and an S-Model using Transformer for network feature sets.
• Investigated the impact of fusion strategies at various interaction depths for enhanced detection.

Main Results:

• Achieved multi-class accuracy rates of 93.40% on the CICIDS2017 dataset.
• Achieved multi-class accuracy rates of 98.50% on the CICIDS2018 dataset.
• Demonstrated superior performance compared to existing few-shot network intrusion detection methods.

Conclusions:

• The multimodal fusion approach significantly enhances few-shot network intrusion detection capabilities.
• Integrating traffic feature graphs and network feature sets effectively addresses data scarcity for rare attacks.
• The proposed G-Model and S-Model fusion strategy offers a robust solution for advanced network security.