- Peter Wonah Odey, Hashim Abdul Isah* and Okoduwa Ernest Atama
- DOI: 10.5281/zenodo.21737248
- SSR Journal of Artificial Intelligence (SSRJAI)
Machine learning-based intrusion detection
systems for IoT networks are predominantly evaluated on detection metrics
alone; computational efficiency is left unmeasured and deployment feasibility
unaddressed. This study presents a dual-axis evaluation of five lightweight ML
models on the TON_IoT network traffic dataset under a bias-aware preprocessing
protocol that removes testbed-specific identifiers before feature selection.
The preprocessing protocol is designed to produce feature sets that generalise
beyond the training topology. A hybrid mutual information and recursive feature
elimination pipeline reduced 60 candidate features to 18 behavioural
attributes. Five models were trained and evaluated on the held-out test
partition under CPU-only, single-core inference conditions (Random Forest,
LightGBM, XGBoost, a Pruned MLP, and a 1D-CNN); each was assessed on four
detection metrics (accuracy, macro-F1, false positive rate, AUC-ROC) and three
efficiency dimensions (per-sample latency, serialised model size, peak RAM at
inference). Tree-based models achieved macro-F1 scores of 0.923–0.936 and false
positive rates of 0.003. Neural network models recorded macro-F1 scores of
0.599–0.637; the margin below the tree models (0.287–0.337 points) is
attributable to calibration effects under class imbalance rather than to
architectural limitations. Random Forest recorded the lowest latency (0.017 ms
per sample) and is recommended for gateway-class deployment. The Pruned MLP, at
218.5 KB and 0.070 ms latency, is the only model satisfying all MCU-class
constraints and is recommended for severely resource-constrained endpoints.
When deployment tier thresholds are applied, four models are assigned to
gateway class and one to MCU class. Detection performance and computational efficiency
impose different model rankings; deployment decisions require both axes of
evaluation.
Keywords: Intrusion
detection system, Internet of Things, Edge deployment, Lightweight machine
learning, TON_IoT, Feature selection, Computational efficiency.
