Transfer Learning Approach to IDS on Cloud IoT Devices Using Optimized CNN

Abstract

Data centralization can potentially increase Internet of Things (IoT) usage. The trend is to move IoT devices to a centralized server with higher memory capacity and a more robust management interface. Hence, a larger volume of data will be transmitted, resulting in more network security issues. Cloud IoT offers more advantages for deploying and managing IoT systems through minimizing response delays, optimal latency, and effective network load distribution. As a result, sophisticated network attack strategies are deployed to leverage the vulnerabilities in the extensive network space and exploit user information. Several attempts have been made to provide network intrusion detection systems (IDS) to the cloud IoT interface using machine learning and deep learning approaches on dedicated IDS datasets. This paper proposes a transfer learning IDS based on the Convolutional Neural Network (CNN) architecture that has shown excellent results on image classification. We use five pre-trained CNN models, including VGG16, VGG19, Inception, MobileNet, and EfficientNets, to train on two selected datasets: CIC-IDS2017 and CSE-CICIDS2018. Before the training, we carry out preprocessing, imbalance treatment, dimensionality reduction, and conversion of the feature vector into images suitable for the CNN architecture using Quantile Transformer. Three best-performing models (InceptionV3, MobileNetV3Small, and EfficientNetV2B0) are selected to develop an ensemble model called efficient-lightweight ensemble transfer learning (ELETL-IDS) using the model averaging approach. On evaluation, the findings show that the ELETL-IDS outperformed existing state-of-the-art proposals in all evaluation metrics, reaching 100% in accuracy, precision, recall, and F-score. We use Matthew's Correlation Coefficient (MCC) to validate this result and compared it to the AUC-ROC, which maintained an exact value of 0.9996. To this end, our proposed model is lightweight, efficient, and reliable enough to be deployed in cloud IoT systems for intrusion detection.

A Survey of Deep Learning Technologies for Intrusion Detection in Internet of Things

“…Standard techniques for extracting features in deep learning include CNN, AE, and DBN. CNN can capture local features in data by performing convolutional operations, which is particularly useful for processing network traffic data [84]. In contrast, AE is an unsupervised learning method that extracts valuable feature representations from unlabeled data.…”

An Attention-Based Convolutional Neural Network for Intrusion Detection Model

“…As a result of feeding the resulting NetFlow images into the designed CNN model, 95.86% accuracy is achieved in detecting the intrusion. OKEY et al [27] trained five pretrained CNNs using the CIC-IDS2017 and CSE-CICIDS2018 datasets, including VGG16, VGG19, MobileNet, Inception, and EfficientNets. After a series of data preprocessing processes, three models (InceptionV3, MobileNetV3Small, and EfficientNetV2B0) were selected based on their performance to develop an efficient-lightweight ensemble transfer learning (ELETL-IDS) model.…”

Enhancing Intrusion Detection in Industrial Internet of Things through Automated Preprocessing

“…In 2015, the UNSW-NB15 dataset was developed to aid in the research of intrusion detection, and a study [7] has utilized this dataset. The main objective of this study is to utilize machine learning techniques to identify significant features that can enhance intrusion detection in network systems, while also addressing the curse of high dimensionality caused by irrelevant and redundant features.…”

“…1,2,3,4,5,7,8,9,11,12,14,15,16,17,20,23, 28, 29, 30, 31, 32, 34, 35, 36, 38, 45, 47, 49, 52, 54, 55, 57, 58, 59 0.9975 2, 3, 4, 5, 7, 11, 14, 16, 17, 18, 21, 23, 25, 27, 29, 30, 31, 32, 33, 34, 36, 39, 41, 43, 44, 45, 48, 49, 50, 52, 53, 55, 57, 58, 59 0.9977 1, 2, 3, 4, 7, 11, 15, 16, 17, 18, 19, 21, 22, 25, 28, 29, 30, 33, 34, 38, 41, 43, 44, 48, 49, 50, 52, 53, 55, 57, 58, 59 0.9982…”