Improving Intrusion Detection in IoT Networks

Networks of interconnected devices exchanging data are becoming popular in several contexts. These Internet of Things (IoTs) range from medical and wearable devices to smart components in industrial and urban infrastructures. Their proliferation increases the surface area for cyber attacks that aim to data pollution and leakage or sabotages blocking systems and possibly asking for ransoms. In many contexts, detection through machine learning remains an open issue because of too many false alarms. They waste valuable time of cybersecurity teams and erode trust in the entire security infrastructure. We think that more attention should be moved from algorithms, that are now mature, to feature extraction (FE) methods. We aim to improve accuracy and precision of detection through a combination of methods. We propose a feature extraction and pre-processing pipeline that provides the model with a high-quality datasets. Moreover, we explore a novel application of the Bayes Point Machine algorithm which, to our knowledge, has never been applied to Network Intrusion Detection Systems. This model offers several advantages, including online learning capabilities, resistance to outliers, and automatic adaptability to feature scaling. We evaluate the proposed solution in a network of IoT-based devices and we can show that the Bayesian model together with the proposed FE offers superior performance.