A Dual-Stage State-Space Attention Framework for Real-Time DDoS Attack Detection on the CIC-DDoS۲۰۱۹ Dataset

The escalating frequency and sophistication of Distributed Denial-of-Service (DDoS) attacks continue to threaten the stability of modern network infrastructures. While deep learning has substantially improved detection capabilities, most existing solutions struggle to balance high classification accuracy with real-time computational efficiency-a trade-off that becomes increasingly problematic in high-throughput and resource-constrained environments. This paper introduces a Dual-Stage State-Space Attention (DSSA) framework that addresses this limitation by synergistically combining a Mamba-based state-space encoder with a sparse Transformer attention module. The proposed architecture processes network traffic sequences with linear-time complexity, capturing long-range temporal dependencies while maintaining an exceptionally low inference footprint. Additionally, a dynamic feature gating mechanism adaptively selects the most discriminative traffic features, further accelerating processing without compromising detection performance. Extensive experiments on the CIC-DDOS۲۰۱۹ dataset demonstrate that the DSSA framework achieves an accuracy of ۰.۹۹۹۹, precision of ۰.۹۹۹۸, recall of ۰.۹۹۹۹, and Fl-score of ۰.۹۹۹۹, with an average inference latency of merely ۰.۰۵۸ milliseconds and a model size of approximately ۸.۲ kilobytes. These results outperform conventional CNN-based, Transformer-based, and existing state-space models, establishing the DSSA framework as a highly effective and deployable solution for real-world intrusion detection systems

کلیدواژه ها:

Network Security ، DDoS Detection ، State-Space Models ، Real-Time Detection ، Deep Learning