Intelligent Network Intrusion Detection Using FEA-CASSTFormerNet Optimized by Archerfish Hunting Algorithm

Abstract

Through the rapid growth in complexity of cyber-attacks on networked platforms, there is a need for advanced and adaptable Intrusion Detection Systems (IDSs) featuring real-time analysis and reliable categorization capabilities. This paper presents an enhanced hybrid approach to IDS based on sophisticated Deep Learning (DL) architectures. It discusses with some of the procedures which involves the following stages. First, the process begins by collecting all of the unprocessed data collected from network log records. Next, each data field goes through an extensive pre-processing phase that addresses not only missing values via a KNN Imputer,  also includes Min-Max scaling to normalize the data as well as Hybrid encoding of categorical variables (one-hot encoding and entity embedding) to create appropriately formatte d data by the DL algorithms. Further analysis of class imbalances is conducted using Exploratory Data Analysis (EDA) methods that takes both distributional data through pie-chart views and correlation assessments comparing all of the different fields of the dataset using a heatmap format. Features are extracted using Shannon Entropy for purposes of meaningful representation learning, while dimensionality reduction and covariance reduction will be accomplished through the use of Principal Component Analysis (PCA). To ensure a fair evaluation, the dataset used for testing is split in half, one for training and the other for testing. The framework suggested is FEA-CasSTFormerNet, this framework combines both recursive attention through memory mechanisms using Fractal Echo Attention (FEA) and deep temporal and spatial relationship modeling using Cascaded Spatio Temporal Transform (CasSTFormer). To balance exploration versus exploitation, the framework utilizes Archerfish Hunting Optimization (AfHO) to adjust hyper-parameters during the archerfish's behavior of chasing prey. Through experimentation with benchmark datasets for IDS inscribed in Python, results demonstrate that the proposed framework provides extremely high accuracy of 0.9998, precision of 0.9988, Recall of 0.9960 and FI-Score of 0.9974 at identifying diverse attack classes, less false positives, and greater generalization ability than existing solutions.