Hierarchical Attention-Aligned Transformer With Multimodal Fusion for Explainable Network Intrusion Detection

With the increasing sophistication of cyberattack techniques, traditional intrusion detection systems face dual challenges of insufficient feature extraction capability and poor decision interpretability. This paper proposes a hybrid Transformer-based detection framework integrated with deep interpretability analysis, which significantly enhances detection performance and decision transparency through optimized data processing pipelines, novel network architecture design, and embedded model interpretation techniques. The approach employs robust normalization and quantile truncation for highdimensional feature processing, designs a parallel Transformer module with multi-head attention to capture spatiotemporal patterns in network traffic, and incorporates LIME interpreters to construct a visual analytics system. Experiments on the CICIDS2017 dataset demonstrate that the model achieves notably higher accuracy than conventional methods while effectively tracing the rationale behind attack decisions. This study provides a new technical pathway for intelligent security analytics systems.