Abstract As digital infrastructure expands, the threat of network intrusions grows, especially from interconnected networks like the Internet of Things and cloud computing. Reliable intrusion detection systems are essential. Deep learning excels at identifying complex patterns but struggles with overfitting and gradient vanishing, limiting effectiveness. To address these challenges, a recent study proposes a framework called GuardianNet, combining an unsupervised layer-wise deep autoencoder with a bidirectional long short-term memory network optimized through hyperparameter tuning processes. This approach improves the ability of the intrusion detection system to extract vital features from raw data, thereby enhancing its ability to distinguish attacks from benign connections. The effectiveness of the proposed approach was evaluated on three benchmark datasets called KDDCUP99, UNSW-NB15, and CICIDS17. Two experiments were conducted on these datasets to justify the methods adopted in GuardianNet. Finally, a comparison is employed between several state-of-the-art methodologies and ours regarding frequent evaluation criteria in the domain.
Fatahnaie et al. (Mon,) studied this question.