IoT Network Intrusion Detection
Five machine-learning classifiers trained on CICIoT2023 (508K instances, 34 attack types), with SHAP feature attribution and noise-robustness tests.
Five classifiers were trained on the CICIoT2023 dataset (508K instances, 34 attack types). The best result reaches 93.22% accuracy on the 34-class task.
SHAP TreeExplainer attribution shows that packet inter-arrival time dominates the models’ decisions. The evaluation includes an ablation study and Gaussian-noise robustness tests.