A BMO-Optimized Bi-LSTM Framework for Robust Intrusion Detection in IoT and Network Environments
Аннотация
More powerful intrusion detection systems are needed due to the fact that cybersecurity threats have gotten smarter as the number of linked devices besides systems has grown exponentially. To achieve precise categorization, this research introduces a new framework that combines the barnacle mating optimizer (BMO) with bidirectional long short-term memory (Bi-LSTM) networks. For both binary and multiclass classification tasks, the suggested model is tested on two well-known datasets: NSL-KDD and BoT-IoT. A thorough preprocessing pipeline is used to guarantee balanced and clean inputs. This pipeline handles missing data, encodes categorical features, and performs class balancing using the synthetic minority oversampling technique (SMOTE) for NSL-KDD and adaptive synthetic sampling (ADASYN) for BoT-IoT. BMO improves model efficiency and decreases overfitting by efficiently lowering dimensionality—53.65% for NSL-KDD and 51.72% for BoT-IoT—by picking the most useful features. The improved classification performance is a result of Bi-LSTM's ability to capture bidirectional temporal relationships. The experimental findings validate that the suggested model attains a binary classification accuracy of 99.46% on BoT-IoT and 98.12% on NSL-KDD. It achieves an accuracy of 96.43% for one-class jobs and 97.98% for another. When compared to both conventional and deep learning baselines, these outcomes perform far better on a number of evaluation metrics, including Precision, Recall, F1-Score, besides AUC-ROC. An important step forward in intelligent intrusion detection for dynamic and complicated network settings, the model excels in both generalization and detection.