Abstract—Intrusion Detection Systems (IDS) are essential for securing enterprise and IoT networks against evolving cyber threats. This study proposes a Machine Learning (ML)-based IDS framework that integrates multiple algorithms to improve detection accuracy and resilience. Using the UNSW-NB15 dataset, models including Decision Tree (DT), Random Forest (RF), CatBoost, and hybrid approaches were trained and evaluated for binary classification of network activities. To mitigate performance degradation caused by high-dimensional feature vectors, a Gini Impurity-Based Weighted Random Forest (GIWRF) was employed for feature selection, while Genetic Algorithm (GA)-based feature extraction further enhanced the model’s understanding of class distributions. A total of twenty-seven features were selected based on their relevance, optimizing the learning process. Experimental results demonstrate that the hybrid model outperforms individual algorithms, achieving high accuracy in detecting various attacks, including DoS, Probe, and other network intrusions. The proposed GIWRF-Hybrid approach showed superior performance in both accuracy and loss metrics, confirming its practical applicability for real-world IoT security scenarios. The study provides insights into the design of robust ML-based IDS frameworks and underscores the importance of customized strategies and continuous improvements to enhance system resilience against increasingly sophisticated cyber-attacks. These findings contribute to strengthening IoT network defenses by combining feature selection, extraction, and hybrid classification methods within a single integrated approach.