Accuracy Alone Is Misleading: A Resource-Aware Benchmark of Intrusion Detection Models on CICIoT2023
A leakage-conscious comparison of eight model families on 46.7 million IoT flows. Quality, rare-class behaviour and deployment cost are evaluated together.
Inside the methodology
Combined multi-seed analysis, paired bootstrap tests, SMOTE/class-weighting ablations, feature selection, calibration, and deployment profiling (latency, throughput, memory, model size, ONNX).
Under the study's evaluation protocol.
