A Hybrid CNN Approach for Unknown Attack Detection in Edge-Based IoT Networks

INTRODUCTION: In the constantly growing Internet of Things (IoT), device security is crucial. As IoT gadgets pervade our lives, detecting unforeseen assaults is crucial to protecting them. Behavioral analysis, machine learning, and collaborative intelligence may be needed to protect against new dangers. This short discusses the need of detecting unexpected IoT attacks and essential security strategies for these interconnected environments. OBJECTIVES: This research uses the BoT-IoT dataset to create an enhanced IoT intrusion detection system. The goals are to optimize a CNN architecture for effective pattern recognition, address imbalanced data, and evaluate model performance using precision, recall, F1-score, and AUC-ROC measures. Improving IoT ecosystem reliability and security against unknown assaults is the ultimate goal. METHODS: we proposed Crow Search Optimizer Based CNN model with blacklisting table approach to deal with known as well as unknown attacks. For the experiment we used BoT-IoT dataset. This involves tuning a Convolutional Neural Network (CNN) architecture to improve pattern recognition. Oversampling and class weighting address imbalanced data issues, for detecting unknown attack we implement blacklist table in that we use threshold based to prevent flood of attacks and use protocol-based attack prevention mechanism. RESULTS: The comprehensive evaluation of our innovative unknown attack detection method shows promise, suggesting it may be better than existing methods. A high accuracy, precision, recall, and f-measure of 98.23% were attained using an advanced model and feature selection methods. This achievement was achieved by using features designed to identify unknown attacks in the dataset, proving the proposed methodology works. CONCLUSION: This research presents an improved IoT Intrusion Detection System using the BoT-IoT dataset. The optimised Convolutional Neural Network architecture and imbalanced data handling approaches achieved 98.23% accuracy. © 2024 R. R. Papalkar et al. All rights reserved.