Abstract
Machine-learning network intrusion detectors fail when moved to a new network, and recent work reports that fine-tuning a source-trained model on a few labelled target flows restores detection. These studies do not train on the same target labels alone, so they cannot tell whether the recovery comes from the source. We test this on the four NetFlow v3 benchmarks (NF-CICIDS2018, NF-UNSW-NB15, NF-ToN-IoT and NF-BoT-IoT) with a leak-free protocol: 12 directed dataset pairs, random forest, gradient-boosted tree and neural-network models, and pooled, importance-weighted and fine-tuned adaptation, each compared with target-only training over 10 seeds with paired Wilcoxon tests. Across 540 comparisons, the source significantly helps in 8% and hurts in 38% (uncorrected). Averaged over pairs it lowers F1 by 0.044 (95% CI −0.066 to −0.022), and the mean is negative for every source and target dataset; on balanced accuracy the effect has the same direction but is smaller. Benefit is rare and pair-dependent. Zero-shot transfer is weak, a published zero-shot F1 of 0.81 does not reproduce on v3, and at a 5% label budget target-only training matches fine-tuning. A domain classifier separates every pair of networks with AUC of at least 0.996 on raw features, even within the same attack category, and 23–40 of 45 features must be removed before this falls below 0.9. Removing the documented TTL artifact does not change the conclusion. We also show that per-domain feature scaling corrupts separability measurements. Cross-dataset studies should report target-only training at the same label budget.