No universal winner in quantum-inspired intrusion detection: the top model changes with the traffic dataset
Synopsis
Quantum-inspired intrusion detectors are often shown through a single high score, leaving open whether that conclusion survives a different split, preprocessing pipeline or deployment budget. Q-IDS Bench evaluates nine quantum-inspired configurations on five network and IoT datasets under one leakage-aware protocol with five random seeds, all on classical hardware. Best weighted F1 ranges from 98.86% to 99.95%, but the leading architecture changes with the dataset, and only about 1.13 percentage points separate the first- and ninth-ranked configurations by five-dataset mean.
Interpretation
Intrusion-detection rankings shift with the traffic dataset, and no single configuration leads on all five. Such quantum-inspired detectors have mostly been presented through a top score on one dataset, which makes it hard to tell whether the conclusion holds under another split or preprocessing pipeline. QRandomForest leads CIC-IDS-2017 at 99.01% weighted F1, QSVM reaches 99.95% on NSL-KDD, QCNN Standalone reaches 99.34% on UNSW-NB15, and QXGBoost leads CICIoT2023 and CSE-CIC-IDS2018 at 99.29% and 98.86%.
Averaged over the five datasets, the nine configurations sit close together, so model choice looks more like an engineering trade-off than a clear statistical winner. A high score on one dataset hides how close the configurations are once cross-dataset means are compared. QSVM ranks first with 98.38% mean weighted F1, followed by Quantum LSTM+Attention at 98.21%, QXGBoost at 98.12% and QCNN Standalone at 98.11%, with only about 1.13 percentage points between the first and ninth means.
Each model fits its preprocessing on the training partition and is then evaluated on held-out test data, with the same protocol across every model-dataset combination. Fitting scaling, feature selection or PCA on the full dataset would let test-distribution information shape the learned representation. Missing-value handling, categorical encoding, scaling, feature selection and PCA are fitted on the training partition only and applied unchanged to validation and test data; every experiment repeats seeds 42, 123, 2024, 2025 and 777, neural configurations train for ten epochs, and checkpoints or thresholds are chosen on validation data.
The configuration with the highest predictive score is not necessarily the strongest on false-positive rate and inference latency. Ranking by F1 alone does not capture the false-alarm and response-time requirements of deployment. The best NSL-KDD configuration reaches 99.95% weighted F1 at 0.0802 ms/sample model-side latency, while the best CICIoT2023 configuration reaches 99.29% at 0.0046 ms/sample; these are model-side inference times, not end-to-end network response times.
Perspective
The evaluation covers nine quantum-inspired configurations on five network and IoT datasets, all executed on classical hardware with no qubits, quantum backend or hardware circuit execution; reported latency is model-side inference only and excludes data capture, feature extraction, transport and alert response.
The authors state that the current models are quantum-inspired classical implementations, and that genuine quantum machine learning still needs a separate evaluation reporting qubit count, encoding, ansatz depth, shots, backend and noise model; the five datasets differ in age, generation process, feature schema and split strategy, so strong within-dataset results do not demonstrate cross-domain transfer; recurrent-style models operate on tabular vectors rather than true host or session sequences; and confusion-matrix counts were reconstructed from aggregate rates rather than exported from stored per-sample predictions.