Intrusion Detection Using CTGAN and Lightweight Neural Network for Internet of Things

Abstract

Deep learning-based intrusion detection systems have high accuracy and low false alarm rates. However, there are challenges to deploy deep learning models in the vulnerable, resource-constrained Internet of Things. Therefore, two deep learning models are proposed: Lightweight Intrusion Detection System using Feedforward Neural Network (LIDSuFNN) and Lightweight Intrusion Detection System using Convolutional Neural Network (LIDSuCNN). In the models, the feedforward neural network is compressed using neuron pruning and the convolutional neural network is compressed using filter pruning. Then, quantization has been applied to the models. The models are trained and tested on standard datasets and synthetic datasets. A generative artificial intelligence model, Conditional Tabular Generative Adversarial Network (CTGAN), has been used to generate synthetic data. The models have been compared with the baselines and results are analyzed. Experimental results show that the proposed models require less training time and memory than the baselines, with approximately similar performance. The reduction of various parameters is due to the fact that pruning and quantization have removed unnecessary calculations from the networks. Statistical analysis has also been done to show the superiority of the proposed techniques.