Abstract:
This research investigates the application of Gradient-based Meta-learning (GBML) and Model-agnostic Meta-learning (MAML) techniques for improving zero-day attack detection in Network Intrusion Detection Systems (NIDS). Traditional methods using labeled data are ineffective against novel attacks exploiting previously unknown vulnerabilities. Our study demonstrates how meta-learning approaches can improve adaptability and detection accuracy for zero-day attacks. The proposed framework utilizes the NF-UQ-NIDS dataset and develops meta-learning models for both anomaly detection and zero-day attack classification. The models were trained using 34 network traffic features including IP addresses, port numbers, protocol identifiers, packet metrics, flow duration, throughput values, and TCP/DNS parameters. Experimental results demonstrate that the GBML framework outperforms MAML significantly. For zero-day attack detection, GBML achieved 86.15% accuracy, 84.32% precision, and 85.67% recall, compared to MAML's 67.24% accuracy, 65.48% precision, and 66.91% recall. For normal traffic detection, GBML attained 85.08% accuracy with 83.75% precision and 84.26% recall. These results indicate that GBML provides better generalization and efficiency for dynamic, real-world cybersecurity applications. This research contributes to the growing body of knowledge in applying advanced machine learning techniques to combat evolving cyber threats, highlighting the need for scalable and adaptive solutions to protect network infrastructures.
