SPC_BiGLNet - An efficient Intrusion Detection technique using Intelligent Feature Selection

What the paper says

The Internet and communication have developed very fast, and an expanding network and size of information have been achieved. Threats created by attacks are really very challenging to network security. Subsequently, intrusion detection technology is intended to identify network attacks besides ensuring data privacy, accessibility, and integrity. There are a number of methods that were designed to find intrusions by a network. But within the process of intrusion detection, there have been some constraints experienced, such as poor performance, overfitting, few detections of attacks, etc. In order to eliminate these problems, the proposed methodology will develop an efficient Intrusion Detection System (IDS). The first step is to extract data via a dataset, and data augmentation is done under the K-Nearest Neighbor (KNN) compression in the data sampling process. The Hybrid Giant Tent Armadillo with Active Weight Kookaburra Optimization (HGTA2WKO) algorithm is then applied to best feature selection. Lastly, intrusion detection is carried out with the help of the new Stacked Primary capsule-assisted Bidirectional Gated Lyrebird Network (SPC-BiGLNet) model. In order to establish the efficiency of the proposed technique, the performance of the technique is compared with that of some related techniques. The two datasets, the UNSWNB-15 dataset and the CICIDS-2017 dataset, were utilized in this study for the detection of intrusion. The two benchmark datasets that are used to evaluate the model include CICIDS-2017 and UNSWNB-15. The accuracy, Precision, Recall, F1-score, and False Alarm Rate of the proposed SPC-BiGLNet on the CICIDS-2017 dataset are 99.74, 99.74, 99.74, 99.74 and 0.00256, respectively. It has an Accuracy of 99.41, Precision of 99.37, Recall of 99.42, F1-score of 99.40 and FAR of 0.0057 on the UNSWNB-15 dataset. Exploring the TON-IoT data, the model shows 99.53% accuracy, 99.50% precision, 99.48% recall, and 99.49% F1-score, which proves its efficiency in the case of contemporary network traffic. The confusion matrix analysis also provides further recommendations that the model is very specific, with little misclassification of normal and attack cases. Compared to prevailing systems such as CNN-BiLSTM-Attention, SAE, DBN, DCRNN, and RNN, the proposed method consistently outperforms in all performance metrics. These results validate the effectiveness of the proposed HGTA2WKO-based feature selection and the SPC_BiGLNet detection framework in building a highly accurate, low-latency, and generalizable intrusion detection system.