Machine Learning-Based IoT Intrusion Detection: A Survey and Performance Comparison

As IoT devices multiply, the cyber attack surface has widened just as fast, and with it the need for intrusion detection systems (IDS) that are efficient, dependable, and protective of user privacy. This paper organizes IoT IDS methods into a taxonomy built around training architecture and reviews them under two main paradigms, misuse detection and anomaly detection, with the focus on centralized traditional machine learning and deep learning. Experiments on several public datasets point to a clear pattern: in centralized settings, tree-based ensemble models and the hybrid CNN-LSTM architecture give the strongest overall results, whereas Isolation Forest holds up better once the data is compressed at the edge. For federated deployment, tree-based ensembles stand out as the preferred local classifiers.