Improving IoT Security by A Hybrid BiLSTM-WOA Approach for Robust Attack Detection

Abstract

The rapid development of IoT devices has attracted the attention of researchers and created significant new security challenges. This challenge has highlighted the need for advanced methods and techniques to detect and address intrusions. This paper presents a robust and coherent system for detecting IoT attacks using the hybrid BiLSTM-WOA model. Our research uses two large datasets, CIC IoT 2023 and N-BaIoT, which contain 46,686,579 and 863,057 records covering multiple IoT attack scenarios, respectively.  An important difficulty tackled in this work was the noticeable data imbalance in both datasets, where attacks were frequent compared to normal traffic. This was resolved by implementing meticulous preprocessing techniques that allowed us to balance the datasets to achieve a 50-50 distribution between attack and regular traffic. This led to unbiased model training and evaluation of the trained models. The innovative integration of the Whale Optimization Algorithm (WOA) with the BiLSTM model enabled the automated fine-tuning of essential hyperparameters, precisely the number of LSTM units and dropout rate. This led to improved model generalization and performance. We conducted a thorough performance evaluation using accuracy, precision, recall, and F1-score metrics. The BiLSTM-WOA model achieved an impressive accuracy rate of 99%, surpassing other LSTM variants; Vanilla LSTM and Time-Distributed LSTM achieved 97%, while Deep LSTM and Stacked LSTM had accuracies of only 40% and 47%, respectively. Solutions developed with training processes showed a general improvement in accuracy with a slight overfitting trend due to the close alignment of curves for training and validation accuracy. These findings confirm the strength of the model and its applicability in the practical implementation of IoT security. This analysis may contribute to IoT security by providing a potent new model incorporating various fusion schemes and algorithms within its framework. It provides a precise, highly reliable and scalable attack model.