The growing complexity and evolving nature of cyber threats and the needs to detect such threats across different network environments, IDS rely more and more on machine learning (ML) models. In this systematic review, we review the most recent ML-based IDS with emphasis on the ML model type, the ML model performance metrics, challenges in ML model development, and emerging trends in ML model design for IDS. We find that models like Neural Networks and Random Forests achieve high accuracy and adaptability, but are not practical due to real-time processing challenges, data privacy, limited datasets, and interpretability issues. Federated learning (FL), Explainable AI (XAI), and AutoML are emerging solutions to improve the IDS adaptability, transparency and efficiency. Fillings these research gaps is critical to the practical, ethical, and resilient deployment of ML driven IDS in today’s cybersecurity landscape. These findings are reviewed here to create a roadmap for future research in developing IDS capabilities, highlighting the need for diverse datasets, privacy-preserving techniques, and standardized ethical frameworks.