Optimizing ML classifiers for superior intrusion detection in resource-constrained smart homes

Abstract

Machine learning (ML) has indeed become essential to the enhancement of intrusion detection systems in different scenarios. It has become a critical barrier against various sophisticated cyber threats. Security vulnerabilities pose special challenges to smart homes, given that devices, sensors, and network connections make the ecosystem highly connected. Such systems improve convenience and efficiency but are generally based on hardware with limited processing power and storage capacity. Therefore, these are prone to a variety of potential attacks. To do so, an effective IDS would have to identify known and evolving threats at all the various vulnerable points, starting from network interfaces down to individual devices. This work tackles these challenges by designing and optimizing ML models that offer reliable intrusion detection tailored for resource-constrained smart home environments. This work argues for intrusion prediction in smart homes using the Extra Tree Classification (ETC) and Linear Discriminant Analysis Classification (LDAC). To strengthen these base models' predictive capability, this paper considered the use of 2 optimization algorithms: the Rider Optimization Algorithm (ROA) and the Aquila Optimizer (AO). The optimizers were integrated strategically with the base models for improved accuracy, thus giving rise to new hybrid models. The combination of ETC with AO provides the ETAO model, while ETC with ROA gives the ETRO model. In equal measure, the LDAC model combined with ROA gives the LDRO model, while that of the LDAC model combined with AO gives the LDAO model. Basically, these hybrid models aim to ensure better performance from a prediction perspective. In the test section, the ETAO model was head and shoulders above the others in this metric, with an excellent value of 0.984, while for the ETRO model, the second-best performing model achieved 0.975. Later on, looking at the entire section, the precision metric again scored highest with the ETAO model at 0.987, while the weakest performance was from the LDAO model, which had a value of 0.888.