Adaptive Network Threat Detection Using Attention-Based LSTM

The evolution of the Internet and its pervasive applicability across domains has several security concerns. Network security is one of the important aspects that needs more focus in designing frameworks to mitigate ever-evolving attack landscapes posed by Internet use. Recent studies suggest applying AI tools and techniques to improve cybersecurity frameworks for Intrusion Detection Systems (IDS). This paper proposes an IDS framework by applying the multi-headed attention (MHA) model to long-short-term memory (LSTM) to detect any network anomaly accurately. The performance comparison of the LSTM framework with and without the attention mechanism shows the significance of applying the attention mechanism. This paper thoroughly analyzed the framework on different benchmark datasets to prove its reliability and compared it with several recent state-of-the-art models. The proposed model outperforms many of the existing frameworks, while for a few attack datasets, the performance is close to that of the existing ones. The proposed framework shows a performance of 98.55% on UNSW NB-15, 97.96% on UNSW BoT-IoT, 81.87% on CTU-13, and 96.72% on AWID. The performance across the datasets shows the robustness and reliability of the proposed model.