Answers are grounded in this paper and cite their sources.
Deep models report accuracies at or above 99% on the 5G-NIDD benchmark, yet much of that headline reflects how the data are prepared. We show that two forms of leakage (a random split that mixes capture sessions, and identifier-like features that stand in for the label) inflate results, rebuild the protocol so neither can help, and rerun the reproducible baselines under both protocols. On the clean benchmark we introduce LEAF, a lightweight per-flow attention model whose feature gate and attention pool double as inherent explanations and which performs no intra-window mixing because every mixer we tried lowers macro-F1. LEAF is first on accuracy (98.99%), macro-accuracy, and Matthews correlation coefficient (0.985), and statistically second by a small margin on macro-F1 (97.92%). It comes with inherent explanations about 1800× faster than SHapley Additive exPlanations (SHAP), federated training that stays near 99% on non-identically distributed clients and resists Byzantine poisoning to 25% malicious clients with robust aggregation, and lossless 8-bit integer quantization that cuts memory threefold, to about 29 KB, with sub-millisecond inference. We also report where LEAF loses, namely on cross-extractor transfer, on a novel control-plane attack, and under gradient input perturbations.
No takes yet. Share an insight, caveat, or question.
Hameed et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: