Multi-scale attention fusion for enhanced transformer models in intrusion detection systems

Modern network environments generate vast streams of complex and evolving traffic data, presenting significant challenges for accurate and real-time intrusion detection. Conventional deep learning approaches often fail to effectively capture the multi-scale temporal patterns and are frequently hampered by severe class imbalance and catastrophic forgetting when faced with non-stationary data streams. To overcome these limitations, this paper introduces a novel Multi-scale Attention Fusion (MAF) module, a general-purpose architectural enhancement for transformer-based models designed to achieve synergistic optimization across three critical dimensions: computational efficiency, continual learning adaptability, and multi-scale temporal perception. We present two instantiations of this approach: MEGA+MAF and FNet+MAF . These models synergistically combine short-term local context, long-range dependencies, and global sequence information through an adaptive, learnable gating mechanism. A comprehensive evaluation across four diverse benchmark datasets— ToN-IoT, X-IIoTID, CICEVS2024 , and CSE-CIC-IDS2018 —demonstrates state-of-the-art performance with balanced optimization across all three dimensions: (1) FNet+MAF achieved superior computational efficiency with up to 8.5 × lower memory and 5.3 × faster inference while maintaining high accuracy; (2) MEGA+MAF demonstrated exceptional continual learning capability, achieving 99.10% accuracy in dynamic streaming environments while effectively eliminating backward forgetting ( 0.00% ) and minimizing forward forgetting ( 0.06% ); and (3) Both models exhibited robust multi-scale perception, capturing threats across short-term bursts, mid-range sessions, and global traffic patterns with up to 99.97% F1-score. Our ablation study after 20 epochs of training identifies 80 tokens as optimal, achieving 85.20% accuracy with 145.1 samples/second throughput. Interpretability analyses further confirm that the models learn robust and semantically meaningful feature representations aligned with network security semantics. The proposed framework represents a significant advancement toward building adaptive, next-generation intrusion detection systems capable of evolving with emerging threats while maintaining operational efficiency in resource-constrained environments.