Resisting Quantum Key Distribution Attacks Using Quantum Machine Learning
Paper • 2509.14282 • Published
Quantum-kernel SVC + variational QNN
PCA -> angle embedding -> StronglyEntanglingLayers; fidelity quantum kernel
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 |
0.986667 |
classical_rf_macro_f1 |
0.986629 |
n_classes |
8 |
n_samples |
2400 |
quantum_qnn_accuracy |
0.875 |
quantum_qnn_binary_f1 |
0.933333 |
quantum_qnn_macro_f1 |
0.466667 |
quantum_qnn_n_params |
25 |
quantum_svc_accuracy |
0.908333 |
quantum_svc_binary_f1 |
0.945274 |
quantum_svc_macro_f1 |
0.831611 |
wall_clock_s |
267.1 |
Published baselines this is measured against
- Target metric: macro F1 (synthetic QKD-attack telemetry)
- Baseline: Architecture showcase, not a performance claim. On THIS project's synthetic telemetry neither quantum model beats the classical random forest (RF macro F1 0.987 vs variational QNN 0.467). Published results are MIXED, not uniformly below a random forest -- an earlier revision of this note claimed they were, and the very paper cited here contradicts it: Al-Kuwari et al. (arXiv:2509.14282) Table VIII reports a hybrid QLSTM at 94.7% accuracy / 94.7% F1 against their own random-forest baseline on the same generated data. What remains well-founded is caution about evaluation quality rather than a universal ranking: arXiv:2512.05069 was withdrawn by its own authors for evaluation that was 'insufficient to fully support the conclusions'. That single withdrawal is thin support for a field-wide claim, so the anchor is now a survey: Kaissar, Bou Nassif & Bouridane, Future Internet 18(5):234 (2026-04-27, CC BY 4.0, DOI 10.3390/fi18050234), systematic to January 2026, finds that 'most current approaches rely on simulated quantum environments and legacy datasets, while evaluation practices remain inconsistent across studies' - active but methodologically early, which is a different and better-founded statement than 'worse than a random forest'. The Al-Kuwari paper is now peer-reviewed: IET Quantum Communication 7(1), DOI 10.1049/qtc2.70028 (2026-02-23). And the strongest 2026 result on this exact problem is CLASSICAL, which corroborates the framing here: IACR ePrint 2026/1282 reaches 96.9% recall / 97.6% precision with physics-aware temporal features plus XGBoost, against 17.3% recall for static QBER thresholding.
- Companion metric shown alongside:
classical_rf_macro_f1- reported together because the aggregate figure can look healthy while the class that matters is not.
Training data
- Dataset: QKD attack telemetry (synthetic, methodology-grounded)
- Licence: generated by this repository
- Source: https://arxiv.org/abs/2509.14282 (licence read 2026-08-03)
8 classes = 7 attack types + normal, per Al-Kuwari et al. (IET Quantum Communication 7(1), e70028, 2026). No commercial-OK real QKD-attack capture exists; this is labelled synthetic everywhere it is reported.
Notes and limitations
Shown as an architecture comparison against a random forest on identical data.
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.