Answers are grounded in this paper and cite their sources.
Reported multiclass accuracies for machine-learning intrusion detection on UNSW-NB15 range from roughly 61% to 88%, a spread far wider than any plausible difference between algorithms.This paper separates what the corpus permits from what models achieve.On the official train-to-test partition (175,341 / 82,332 flows), five model families (Random Forest, XGBoost, LightGBM, CatBoost and a hybrid Transformer-Kolmogorov-Arnold Network) are trained under one protocol, with validation-only model selection and five-seed replication.From label conflict among identical feature vectors we derive exact upper bounds that hold for every deterministic classifier on the input each family receives.Label conflict caps per-class F1 at 0.262 for Analysis, 0.271 for Backdoor and 0.679 for DoS; the jointly attainable macro-F1, computed exactly by mixed-integer programming, is 0.785, and the accuracy bound, 93.60%, equals one minus the empirical Bayes error of the test sample.Per-class recall ceilings, by contrast, depend on the objective and differ by a factor of 8.8 for Backdoor.Exact-duplicate conflict accounts for at most 27.7% of the best model's accuracy error and 45.8% of its macro-F1 shortfall.The five families reach 47.5% to 53.0% macro-F1, with Random Forest leading (53.02 ± 0.17), and their errors overlap strongly: 16.1% of test rows are misclassified by all five, against 0.09% expected under independence, and most of these common errors are not forced by label conflict.Five-seed single-factor controls show that a KAN head is indistinguishable from a parameter-matched MLP head, while class balancing lowers accuracy and raises balanced accuracy in every family, tree ensembles included.A random pooled split raises XGBoost accuracy by 6.6 points, but duplicated test rows are classified less accurately than unduplicated ones, so the gain is not memorisation and is best explained by a within-class distribution shift between the official partitions.Applied unchanged to CICIDS2017, the procedure gives an accuracy bound of 99.985%, showing that the severity of label conflict is specific to UNSW-NB15.We recommend reporting per-class ceilings alongside per-class scores.
No takes yet. Share an insight, caveat, or question.
Zahraa Mohammed Hasan (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: