MACML: Marrying attention and convolution-based meta-learning method for few-shot IoT intrusion detection

Area of Science:

• Cybersecurity
• Machine Learning
• Internet of Things (IoT)

Background:

• Internet of Things (IoT) devices are increasingly vulnerable to cyberattacks.
• Traditional intrusion detection systems (IDS) struggle with novel threats due to reliance on extensive labeled data.
• Few-shot learning scenarios present a significant challenge for existing IDS.

Purpose of the Study:

• To develop an advanced intrusion detection method for IoT environments.
• To improve the detection of novel cyberattacks using limited data.
• To enhance the generalization capability of intrusion detection systems.

Main Methods:

• Proposing MACML (Marrying Attention and Convolution-based Meta-Learning), a meta-learning approach.
• Integrating self-attention mechanisms for global feature extraction and convolutional neural networks for local feature extraction.
• Utilizing an optimization-based meta-learning framework for rapid adaptation with few training samples.

Main Results:

• MACML achieved 98.75% accuracy and 99.17% detection rate on CICIDS2018 with only 10 training samples.
• On CICIoT2023, MACML demonstrated 94.47% accuracy and 95.32% detection rate.
• The proposed method outperformed existing state-of-the-art intrusion detection techniques.

Conclusions:

• MACML effectively addresses the challenge of detecting novel cyberattacks in few-shot scenarios for IoT.
• The integration of attention and convolution enhances the model's perception of network traffic characteristics.
• MACML offers a promising solution for robust and adaptable intrusion detection in IoT ecosystems.