Full Length Article
DOI: https://doi.org/10.54216/IJWAC.100107
Time-Forward Evaluation of Denial-of-Service Detection in LEACH-Based Wireless Sensor Networks: Drift, Feature Budgets and Decision Costs
Intrusion detectors for cluster-based wireless sensor networks are scored on shuffled records, although a deployed detector is trained on the early rounds of a network and consulted for the rest of its life. This paper re-examines denial-of-service detection on the WSN-DS benchmark under four protocols, the strictest of which trains on the first part of the simulation and tests on what follows. Five classifiers are compared, class-conditional drift is measured with the Kolmogorov–Smirnov distance, and a feature budget is chosen by forward selection on a validation window. Random and block-grouped folds agree to within half a point of macro-F1. The time-forward split does not: with all fifteen behavioural features a random forest falls from 0.974 to between 0.833 and 0.946 depending on its seed, because most late Blackhole and Grayhole records lie outside the training range. A six-feature forest restores 0.959 with negligible seed variance and cuts the report payload by 60%. Isotonic calibration followed by expectation-maximisation prior correction tracks the rising attack share to within one point and lowers the balanced decision cost by 38%, yet raises it at high miss penalties. Validation scores near the training period could not tell the fragile configuration from the stable one.
Firas Zobary
visibility
1133
download
727