Anomaly-based detection of cyberattacks in smart homes

Smart homes with interconnected IoT devices face growing cybersecurity threats, amplified by remote working, which exposes both personal and corporate data. Although anomaly-based intrusion detection systems are showing promise in the identification of unknown attacks, their reliability is hindered by unrepresentative datasets, severe class imbalance, and inconsistent evaluation practices. This doctoral thesis strengthens the foundations for robust detection systems by investigating the impacts of dataset characteristics, class imbalance, hyperparameter optimization, and metric selection on model effectiveness. Through extensive empirical analysis, this research demonstrates that conventional evaluation fails in imbalanced scenarios. The key finding is that optimizing models using metrics suited for imbalance, such as the Matthews Correlation Coefficient (MCC), yields more reliable and generalizable results. The thesis introduces a validated methodological approach for metric selection to improve hyperparameter tuning, providing actionable guidance for developing effective cybersecurity solutions for real-world smart home environments and addressing current limitations in the field.