Answers are grounded in this paper and cite their sources.
Intrusion detection systems must defend enterprise and IoT networks against increasingly sophisticated threats.We evaluate five machine learning classifiers on 10,000 network flow records covering nine attack categories and a benign class, using 26 features engineered from 13 raw metadata attributes.Departing from common practice, we partition before balancing, so the held-out set of 3,000 flows retains the 24 to 1 class ratio of the source data; hyperparameters are tuned on the training partition alone and standardisation is fitted inside a leakage-free pipeline.Gradient boosting attains the highest area under the precision-recall curve (0.670), detecting 66.67% of attacks at a 1.28% false-positive rate, with random forest close behind (0.659).Evaluation protocol dominates model choice: logistic regression appears competitive on a balanced partition, yet its attack-class precision falls to 13.91% at the 4% prevalence of the source data.Binary detection is viable from raw metadata; fine-grained attribution is not.Because all results come from one simulated distribution whose generator is unavailable to us, they characterise that distribution rather than deployed intrusion detection, and the feature rankings describe the fitted models rather than establishing that the underlying fields are intrinsically diagnostic of attacks.A prevalence-sensitivity analysis shows how precision degrades at the lower priors typical of production traffic.
No takes yet. Share an insight, caveat, or question.
Alzghoul et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: