Energy-efficient intrusion detection with a protocol-aware transformer-spiking hybrid model.

Area of Science:

• Computer Science
• Artificial Intelligence
• Network Security

Background:

• Deep learning and transformer models show promise for intrusion detection but face challenges with computational cost, energy efficiency, and imbalanced data.
• Existing methods struggle with detecting rare attack classes in heterogeneous network traffic patterns.

Purpose of the Study:

• To propose a novel Transformer-Augmented Spiking Neural Network (TASNN) for intrusion detection systems (IDS).
• To enhance robustness to diverse network traffic and improve detection of rare attack classes.
• To reduce computational overhead and increase energy efficiency for edge-based IDS.

Main Methods:

• Developed TASNN integrating attention mechanisms with energy-efficient spiking neural networks.
• Incorporated Protocol-Aware Adaptive Normalization (PAAN) and Pseudo-Flow Reconstruction (PFR) for traffic pattern robustness.
• Utilized adaptive spike encoding (MASE, EDC) and Cross-Modal Gating (XMG) for efficient feature representation and dynamic regulation.
• Employed Spike-Aware Information Fusion (SAIF) for stable and interpretable feature selection.

Main Results:

• TASNN demonstrated improved classification performance on benchmark datasets compared to existing methods.
• The proposed model achieved reduced computational overhead and enhanced energy efficiency.
• TASNN showed superior detection capabilities, particularly for rare attack classes in imbalanced network traffic.

Conclusions:

• TASNN offers a promising solution for intrusion detection, balancing high accuracy with computational and energy efficiency.
• The framework is well-suited for resource-constrained environments and edge-based intrusion detection scenarios.
• The integration of spiking computation and attention mechanisms provides a robust and efficient approach to network security.