An Enhanced Intrusion Detection System Using Attention-Based Stacked Sparse Autoencoder Feature Extraction

Attention-based stacked sparse autoencoders (AB-SSAEs) are an innovative method for improving Intrusion Detection Systems (IDSs) through the extraction of important features in high-dimensional and heterogeneous data. The proposed AB-SSAE presents an innovative approach to optimizing feature extraction processes using attention mechanisms and a hierarchy of focused sparse autoencoders. The AB-SSAE architecture employs several layers of sparse autoencoders, which transform features through attention mechanisms at every level, improving precision for feature extraction. AB-SSAE employs adaptive denoising with median filtering as a preprocessing step. From the mined data, normal and intrusion attempts are efficiently classified using a Bidirectional Long-Short-Term Memory (Bi-LSTM) network. The proposed technique was compared with several existing approaches, and the results showed that it can differentiate between malicious and benign network traffic with an accuracy of over 0.98.