Abstract: The growing volume and sophistication of cyber threats necessitate Network Intrusion Detection Systems (NIDS) that are both accurate and suitable for real-time operation. Existing deep learning models often rely on computationally expensive architectures while overlooking temporal dependencies within network traffic flows. To address this challenge, we propose HALCN, a unified framework for lightweight temporal modeling that integrates multi-scale separable convolutions with low-rank self-attention to jointly capture local and long-range dependencies in network traffic. Unlike existing approaches that prioritize detection performance at the expense of computational efficiency, HALCN is designed to optimize the trade-off between detection accuracy and inference efficiency, enabling real-time deployment in resource-constrained environments. The proposed model incorporates channel attention (ECANet) and Linformer-based temporal self-attention to enhance feature representation while maintaining low latency. To improve robustness under class imbalance, we employ hybrid data augmentation using SMOTE and CGAN, along with an outlier-aware instance selection strategy to reduce redundancy. Evaluated on the CICIDS2017 and CSE-CIC-IDS2018 datasets, HALCN achieves up to 99.83% accuracy with low variance and a low false alarm rate while maintaining low model complexity compared with LSTM-based approaches. These results demonstrate the effectiveness of HALCN as a practical and efficient solution for real-time NIDS deployment.

 

Presenter: Danial Chakma (Std No. 1018052001)

Venue: Graduate Seminar Room