Daemons-Q/qsh-telecom-5g-nidd-gbm

5G-NIDD network-attack classifier (9 classes)

gradient-boosted trees over 47 flow-statistic features

Part of QSMPC-QKD-QHE-AI-Hybrid, a quantum-safe orchestration demo. This is the plaintext model for the telecom_ids use case; the encrypted path runs a distilled student, not this model.

Measured performance

metric	value
aggregate_accuracy	0.982897
leak_counterfactual_accuracy	0.999734
leak_counterfactual_accuracy_gain	0.016837
leak_counterfactual_macro_f1	0.999505
macro_f1	0.995019
majority_baseline	0.392911
n_classes	9
n_classes_predicted	9
n_eval	303973
n_features	47
n_samples	1215890
wall_clock_s	92.6

Published baselines this is measured against

• Target metric: macro F1 across 9 network-attack classes
• Baseline: Two published figures on 5G-NIDD are usually quoted - Wang, Fok & Thing report 99.61% accuracy for a CNN and 99.71% for a hybrid quantum-classical Quan-ConvCNN - and BOTH are accuracies, which is why no number appears in baseline. Three caveats. (1) Those are hybrid quantum-classical results with a 0.10-point spread, and the authors' own claim is about generalisation to UNSEEN attacks rather than about the headline; references/REFERENCES.md WS2 records this. (2) Accuracy is the wrong comparator on a corpus whose two majority classes are 39.3% and 37.6% of the rows. (3) MEASURED HERE: an identical gradient-boosted model reaches accuracy 0.9997 with the corpus's integer row counters restored, so the published range is reproducible from row identifiers alone. ROUND 50 CORRECTED WHAT THAT THIRD POINT WAS TAKEN TO MEAN. It was written when this model scored 0.7610 and read as evidence that the published figures probably leaned on those counters, on the unstated premise that nothing else could close a 24-point gap. Something else could. Refitted on the corpus's 113-column export with every identifier and every host fingerprint dropped, the same configuration scores 0.9829 - about 1.3 points below the published accuracies rather than 24 - and the column that closes the gap is one flow statistic, IdleTime, which the 52-column export does not carry. The leak measurement stands as a fact about row counters. The insinuation about the papers does not, and is withdrawn. These figures are an anchor and not a target that has been beaten or missed.
• Companion metric shown alongside: aggregate_accuracy - reported together because the aggregate figure can look healthy while the class that matters is not.

Gap to the published baseline

There is no baseline number to have a gap to, and that is a deliberate registry decision rather than a missing measurement: the two figures usually quoted for this corpus (99.61% for a CNN, 99.71% for a hybrid quantum-classical Quan-ConvCNN, Wang/Fok/Thing) are ACCURACIES, and this model is judged on macro F1, so putting them in baseline would render an accuracy beside a macro F1 as though the two were the same scale. WHAT THE GAP ACTUALLY IS, now that it has been measured properly. This model scores about 0.9829 accuracy with every row identifier and every host fingerprint dropped, so it sits about 1.3 points below those published accuracies rather than the 24 points it sat below them last round. Against the 9-class multiclass figure that is directly comparable in task - BiTAD, Future Internet 17(11) 482, accuracy 0.9047 - it is above, though on its own split rather than a shared one, so that is an anchor and not a head-to-head. THE CORRECTION THAT MATTERS. Round 49 recorded that restoring the corpus's integer row counters reproduces the published range and implied the published range was therefore explained by them. The first half is still measured true here and is recorded every run as leak_counterfactual_accuracy. The second half does not follow and is withdrawn: 0.9829 is reachable on this corpus with no identifiers at all, using one additional flow statistic (IdleTime) that Combined.csv does not carry, so most of the gap Round 49 attributed to leakage was its own choice of export. What the counterfactual shows is what a row counter is worth, which is a statement about this corpus and not about anyone's paper.

Training data

• Dataset: 5G-NIDD (BTS1_BTS2_fields_preserved.csv)
• Licence: CC-BY-4.0
• Source: https://doi.org/10.23729/e80ac9df-d9fb-47e7-8d0d-01384a415361 (licence read 2026-08-16)
• Attribution (CC-BY-4.0): Y. Siriwardhana et al., 'Descriptor: 5G Wireless Network Intrusion Detection Dataset (5G-NIDD)', in IEEE Data Descriptions, doi:10.1109/IEEEDATA.2025.3592888. Dataset: doi:10.23729/e80ac9df-d9fb-47e7-8d0d-01384a415361, CC BY 4.0.

1,215,890 real 5G network flows x 113 columns, counted in this file rather than quoted, 9 classes: Benign 477,737 | UDPFlood 457,340 | HTTPFlood 140,812 | SlowrateDoS 73,124 | TCPConnectScan 20,052 | SYNScan 20,043 | UDPScan 15,906 | SYNFlood 9,721 | ICMPFlood 1,155. LICENCE VERIFIED AT THE SOURCE, not at a mirror: the Metax record for dataset 9d13ef28-2ca7-44b0-9950-225359afac65 reads access_type Open, license CC-BY-4.0 and restriction_grounds []. The IEEE DataPort listing whose subscription wall this project originally read as a licence restriction is a MIRROR; the Fairdata record is the authors' repository of record. THE EXPORT CHANGED IN ROUND 50 AND THAT IS THE MOST IMPORTANT LINE HERE. This is BTS1_BTS2_fields_preserved.csv, 113 columns. The release also ships Combined.csv at 52 columns, which is what this project trained on until Round 50. Same rows, same classes, same distribution - and all 52 of Combined's columns are informationally identical in the wide file, equal distinct-value counts on every one, so Combined is a strict column SUBSET. It omits IdleTime, and that single column is worth 22 accuracy points: 0.9829 with it, 0.7616 without. Five other columns Combined lacks (Dir, Flgs, TcpOpt, sVlan, dVlan) are worth 0.0006 accuracy between them - which is redundancy rather than emptiness, since Dir alone separates ICMPFlood at one-vs-rest AUC 0.9811 and TcpOpt separates SlowrateDoS at 0.9433. The other 41 columns already carry what they carry. SIXTY-SIX OF THE 113 COLUMNS ARE UNUSABLE and the loader drops them with reasons attached (qsh.models.datasets.NIDD_DROPPED, republished in the model's feature_spec.json): 3 labels (Attack Type, its binary collapse Label, and Attack Tool which scores accuracy 0.9615 on its own), 9 row identifiers, 10 host fingerprints (addresses, MACs, ports, IP ids), 5 columns bit-identical to Dur on every row (RunTime/Mean/Sum/Min/Max), and 39 that hold exactly one state. 47 features remain. THE 39 SINGLE-STATE COLUMNS ARE WORTH A SEPARATE SENTENCE because their names promise signal: 30 of them are 100% NULL in this release, and that includes every SIntPkt/DIntPkt inter-packet-time column and every jitter column. They are empty, not weak. THE ROW IDENTIFIERS REMAIN A FINDING, with a narrower claim than Round 49 made. Unnamed: 0 is the CSV's own index and restoring it alone takes accuracy from 0.9829 to 0.9997. That is what a row counter is worth on this corpus. It is NOT the explanation for the published 99.6-99.7% figures - see the ModelSpec below, which withdraws that.

Notes and limitations

NETWORK-LAYER flow classification on real 5G traffic. This is a DIFFERENT TASK from the QKD-layer attack classifier on this same page, with zero overlap in label space or feature space, so neither score is the other's baseline. REFITTED IN ROUND 50 ON A WIDER EXPORT OF THE SAME CORPUS, which is the fact that explains every number here. It trained on Combined.csv (52 columns) and scored accuracy 0.7608; it now trains on BTS1_BTS2_fields_preserved.csv (113 columns) and scores about 0.9829. The rows are identical and the method is unchanged. One column does all of it: IdleTime, a flow statistic the narrow export omits. Removing it alone from the wide fit returns the model to 0.7616, and that ablation is re-measured on every run as idle_time_ablation. It is not an identifier - 17 distinct integer values over 1.2 million rows - and it separates SlowrateDoS, an attack defined by idling, at one-vs-rest AUC 0.9574. READ THE TWO NUMBERS THE OTHER WAY ROUND FROM EVERYWHERE ELSE IN THIS PROJECT - STILL. Every other model here reports an aggregate that flatters and a per-class figure that does not. On this corpus macro F1 is the OPTIMISTIC number and aggregate accuracy the pessimistic one, and the refit shrank that inversion without removing it: the gap was 0.1655 (0.9263 macro F1 against 0.7608 accuracy) and is about 0.0121 (0.9950 against 0.9829). The mechanism is the same one as before, not a new one - Benign and UDPFlood are 39.3% and 37.6% of the rows and remain the only pair that confuses, so accuracy is dominated by them while the macro average is carried by seven classes at 0.9999 or better. The residual error is 1.71% of the test split against 23.92% before, and it is one direction: UDPFlood predicted as Benign. THE LEAK COUNTERFACTUAL IS STILL RECORDED EVERY RUN and its claim is now narrower. Restoring the corpus's integer row identifiers takes accuracy to about 0.9997, the same shape of result Engelen et al. (WTMC 2021) produced for CIC-IDS2017. Round 49 read that as an explanation for the published 99.6-99.7% figures; it is not, because 0.9829 is reachable here with no identifiers at all -- see M_NIDD's baseline_source, which withdraws that reading.

Honest scope

This model is published as part of a research proof of concept, not as a production system. Numbers above are what this repository measured on the split described, with the code in scripts/train/. Where a figure is carried from the literature rather than measured here, it is labelled as such.