In the face of rapidly growing security threats on the Internet of Things (IoT) networks, machine learning (ML) integration shows promise in identifying cyberattacks. However, while the traditional ML models are effective in certain areas, they often fail to detect complex patterns and unusual behaviour in IoT data due to their difficulty adapting or generalizing. Ensemble learning models utilize the strengths of multiple base models to provide a promising solution but are largely influenced by the choice and proper setting of hyperparameters. This paper explores the impact of hyperparameter tuning on ensemble-based ML models for detecting IoT-related cyberattacks. We conducted a series of experiments utilizing the imbalanced and balanced CICToNIoT datasets, with a focus on binary and multi-class classification. The study assessed the Random Forest (RF) and Extreme Gradient Boosting (XGBoost) models with default hyperparameters after applying three tuning methods: Bayesian Optimization with Tree-structured Parzen Estimators, Grid Search (HGS), and Random Search. Our results reveal that HGS significantly enhanced performance, with XGBoost achieving an accuracy of 99.34% and an F1-score of 99.34% in binary classification, and RF achieving an accuracy of 93.76% and an F1-score of 93.73% in multi-class classification. RF demonstrated strong detection capabilities across various attack types, though it struggled with distinguishing certain attacks. These findings highlight the importance of hyperparameter tuning in enhancing the effectiveness of ML models for IoT cybersecurity.