D 2 SaAN : Hybrid Attention and Learning Based Distributed Deep Self‐Adaptive Autoencoder Network for Intrusion Detection

ABSTRACT With the rapid development of IoT, the networks continue to expand in size, creating the scope for attackers to develop new kinds of attacks to disrupt services. Therefore, network security becomes increasingly crucial to our day‐to‐day interactions for mitigating and detecting malevolent actions within the network environment. Currently, numerous machine learning‐based approaches are introduced to mitigate malicious activities. However, the existing methods produce incompetent results and highly rely on the physical design of traffic features. Hence, a Stacked Hybrid Attention‐based Distributed Deep Self‐adaptive Autoencoder Network (SHA‐D 2 SaAN) model is proposed in this research to overcome existing challenges and detect the intrusions accurately. Specifically, the application of Stacked Hybrid Attention (SHA) mechanism improves the model's capability to recognize the intricate and multi‐stage attacks through capturing the cross‐feature and long‐range dependencies. Moreover, the incorporation of self‐adaptive and drift‐based learning mechanisms improves the model's training efficiency by adaptively tuning its learning rate and aids in enhancing the model's robustness in identifying the network intrusions. The effectiveness of the proposed SHA‐D 2 SaAN model is evaluated using three standardized benchmark datasets, reporting the high accuracy values of 97.7%, 96.9%, and 96.3% over the Edge‐IIoT dataset, the UNSW‐NB15 dataset, and CICIDS dataset, respectively.