ML-based Novelty Detection and Classification of IoT Threats using Network Traffic Analysis

Abstract

This article presents a practical evaluation of machine-learning models to detect novelties and classify threats in IoT networks using an ML-based approach. Given the escalating significance of analyzing network traffic amidst the proliferation of devices and sensitive data exchange, this research holds significant relevance. The IoT Network Intrusion dataset was chosen for experimentation, followed by data processing and imbalance handling techniques. Four distinct models encompassing novelty detection and classification were trained, allowing for an in-depth comparison of their performance in terms of accuracy and time. Notably, after attaining the results, it was evident that these models achieved remarkably high accuracy in novelty detection and classification tasks, emphasizing that techniques based on machine learning can be successfully applied to this context.

A practical evaluation of AutoML tools for binary, multiclass, and multilabel classification

“…"Time" here can have two meanings: "training time", the time taken for the model(s) to fit with the data, and "testing/prediction time", which is the time taken to classify all samples in the test set. This study will not focus on the latter metric since the fast prediction is only crucial in real-time scenarios, such as in Network Traffic Analysis (NTA)-based systems (Aragão et al, 2022(Aragão et al, , 2023. However, the training time is an important analytical aspect since impractical training times can lead to improper (or even unavailable) models.…”