Parallel Local-Global Feature Extraction and Cross-Attention Gated Fusion for Network Intrusion Detection.

With the rapid advancement of network technology and the Internet of Things (IoT), massive, high-dimensional traffic data pose significant challenges to Network Intrusion Detection Systems (NIDS). Existing deep learning methods face two major limitations: (1) Insufficient feature extraction: single models struggle to concurrently capture local and global features, while serial hybrid architectures are prone to feature information loss. (2) Superficial feature fusion: existing strategies overlook deep interactions between features and lack adaptive filtering mechanisms to eliminate redundant information. To address these issues, a Parallel Local-Global Feature Extraction (PLGFE) module is constructed, incorporating a dual-stream architecture with parallel Temporal Convolutional Network (TCN) and Bidirectional Long Short-Term Memory (BiLSTM) branches. This module utilizes TCN's dilated convolutions and BiLSTM's bidirectional memory units to concurrently extract complementary local and global features. The Cross-Attention Gating (CAG) module integrates cross-attention and gating mechanisms for feature fusion: the former enables deep interaction between heterogeneous features, while the latter adaptively selects key features and suppresses redundant interaction features through gated weights. Experiments on the CICIDS2017 and CICIoV2024 datasets demonstrate the proposed method achieves multi-classification accuracies of 99.85% and 99.64%, with F1-Scores of 97.03% and 97.64%, respectively. Comparisons and ablation experiments confirm its competitive detection performance and component effectiveness.