Artificial Intelligence–Driven Intrusion Detection Systems: A Systematic Review of Machine Learning, Deep Learning, and Adaptive Ensemble Approaches

Abstract

This is a critical systematic review of 40 peer-reviewed articles published between 2021 and 2025 on artificial intelligence-enhanced Intrusion Detection Systems (IDS).To maintain rigour and transparency in the review, the PRISMA 2020 and Kitchenham guidelines are adhered to.The appropriate literature was also searched in six major databases, such as IEEE Xplore, ACM Digital Library, ScienceDirect, SpringerLink, Scopus, and Web of Science, and concentrated on the IDS methods, which used machine learning (ML), deep learning (DL), reinforcement learning (RL), and ensemble methods.The results show that there is a clear shift of traditional ML algorithms, including SVM and Random Forest, towards advanced deep learning and hybrid ensemble models that have a higher detection accuracy and flexibility in a variety of benchmark datasets, including CICIDS2017 and UNSW-NB15.The frameworks of reinforcement learning also increase the autonomy of IDS by learning on its own and adapting dynamically, but they have difficulties with the issues of computational efficiency, and the stability of the convergence.On the whole, the review demonstrates that there is ongoing methodological standardization deficit, diversity of datasets, and inter-study reproducibility absence.It finds that explainable, adaptive, and federated AI models should be the main focus of future IDS studies to attain scalable, transparent, and trustworthy cybersecurity frameworks that are able to conduct proactive and intelligent network protection.