LIGHTWEIGHT DEEP LEARNING-BASED INTRUSION DETECTION FOR RESOURCE-CONSTRAINED IOT EDGE NETWORKS

Abstract

With a rising number of connected devices and increasing sophistication of cyber-attacks, the rapid growth of The Internet of Things (IoT) has presented new security challenges considering the limited computational resources of the edge devices and the potential for greater sophistication of attacks. Traditional approaches to protecting against intrusions (IDS) are not always effective in implementing real-time applications for the IoT because of the need to achieve a balance between accuracy and computing efficiency. Considering the highly resource-limited condition at the edge, this study presents a lightweight deep learning framework for real-time intrusion detection in the IoT, specifically targeting resource constrained scenarios at the edge. Four fully connected deep learning architectures with an increasing number of parameters were designed and compared to 2 classical machine learning algorithms, Random Forest (RF) and Logistic Regression (LR). The proposed framework incorporates data preprocessing, feature normalization, training of the model and comprehensive performance evaluation based on various classification and computational metrics. Experimental results show that the accuracy of the Standard DL model is 97.2%, and the Lightweight DL model gives an accuracy of 95.8%, when our model has much lower computational complexity. The Tiny DL model achieved the biggest advantage in terms of model size (0.87 MB), the lowest in terms of trainable parameters (45.7K), and the fastest inference time (3.45 ms), in terms of being deployable on the edge. Comparative analysis reveals that, in resource constrained IoT environments, the proposed Lightweight DL architecture offers the optimum model for predicting the intrusion level while ensuring low latency for ID. The proposed framework has an effective solution to practically implement a next generation edge computing system for execution of IoT tasks.