Abstract
Recent intrusion detection studies have achieved high accuracy using deep learning and transformer-based models; however, many approaches suffer from high computational cost, limited energy efficiency, and poor detection of rare attack classes in imbalanced network traffic. To address these challenges, this study proposes a Transformer-Augmented Spiking Neural Network (TASNN) that integrates attention-driven contextual modeling with energy-efficient spiking computation for intrusion detection systems (IDS). The framework incorporates Protocol-Aware Adaptive Normalization (PAAN) and Pseudo-Flow Reconstruction (PFR) to improve robustness to heterogeneous traffic patterns. An adaptive spike encoding strategy, including Multi-Scale Adaptive Spike Encoding (MASE) and Eventified Delta Coding (EDC), converts tabular features into sparse spiking representations. In addition, a Cross-Modal Gating (XMG) mechanism dynamically regulates spiking activity, while Spike-Aware Information Fusion (SAIF) supports stable and interpretable feature selection. Experimental evaluation on benchmark datasets demonstrates that TASNN achieves improved classification performance and reduced computational overhead compared to existing methods, highlighting its suitability for energy-constrained and edge-based intrusion detection scenarios.