Reproducible framework for the controlled comparison of quantum and classical kernels under distribution shift — the artifact behind the manuscript:
Quantum and Classical Kernels under Distribution Shift: Kernel Geometry Governs Out-of-Distribution Robustness
Within each experimental setting, the classifier, preprocessing, splits, model-selection logic, and bandwidth-tuning freedom are held fixed; only the kernel changes.
- 4 benchmark scenarios, 2 modalities: EMBER (static PE malware), UNSW-NB15 (DoS, Reconnaissance) and ToN-IoT (Scanning) network flows.
-
3 drift mechanisms: within-class sparsity extremes (
$m1$ ), train-geometry extremes ($m2$ ), and natural regime drift on network traffic. - 2 classifier families consuming the same precomputed Gram matrices: SVC and a Laplace-approximation Gaussian process classifier (calibrated uncertainty under shift).
-
11+ kernels: linear, RBF, polynomial, Laplacian, Matérn (median heuristic) + 4 fidelity feature maps — with a symmetric bandwidth sweep (RBF
$\gamma$ factors vs. quantum angle scales). - 72 principal settings × 15 repeated runs, paired Wilcoxon tests with Holm correction.
- Kernel-geometry analysis: effective rank, centered kernel-target alignment (ID/OOD), concentration, and the geometric difference of Huang et al.
- Against the customary linear+RBF baselines, fidelity-based quantum kernels improve OOD balanced accuracy nearly uniformly (69–71 of 72 settings;
$p<10^{-9}$ on network data). - A median-heuristic Laplacian kernel neutralizes the advantage on EMBER at fixed bandwidth — the operative property is geometric, not quantum per se.
- With symmetric bandwidth tuning the advantage re-emerges on EMBER, and under natural network drift the quantum family wins 54/54 settings under the GP classifier (
$p=1.6\times10^{-10}$ ). - One mechanism explains the pattern: OOD accuracy tracks the survival of kernel-target alignment under shift (positive within-setting correlation in 89–99% of 360 setting-seed units, all four datasets), with high effective rank of the training kernel as its structural precondition.
src/
utils/ember/ EMBER export + master/q-split construction
utils/netflow/ network-flow export + shift constructions (m2-centroid, natural)
experiments/ kernel-swap runners (classical, quantum, extended+GPC)
analysis/ kernel-geometry descriptors (eff. rank, KTA, geometric difference)
scripts/
ember/ netflow/ grid drivers (settings x seeds x sizes)
analysis/ family comparisons, Wilcoxon, mechanism tests
reporting/ every table and figure of the paper, generated from results/
results/ frozen aggregated results, tables_v2/, kernel_geometry/, mechanism/
manuscript/ LaTeX source (Springer sn-jnl), figures, cover letter
conda env create -f environment.yml && conda activate kernel-shift-framework
# 1) EMBER grid (masters + q-splits + geometry), then extended kernels + GPC
python scripts/analysis/run_kernel_geometry_grid.py --save-spectra
python scripts/ember/run_extended_kernels_grid.py --qsplit-seeds 42 123 999 7 2024 --model-seeds 42 123 999
# 2) Network-flow grid (exports, shift splits, runners, geometry)
python scripts/netflow/run_netflow_grid.py --qsplit-seeds 42 123 999 7 2024 --model-seeds 42 123 999
# 3) Analyses, tables, and figures of the paper
python scripts/analysis/compare_extended_families.py
python scripts/analysis/mechanism_generalization.py
python scripts/reporting/make_v2_tables.py
python scripts/reporting/make_v2_figures.pyRaw EMBER (2018, feature version 2) must be placed under data/raw/ember/; the network-flow scenarios are exported from the public UNSW-NB15 and ToN-IoT benchmarks (see src/utils/netflow/). A label-leakage sanity check is included (scripts/netflow/check_label_leakage.py).
The manuscript source lives in manuscript/ (Springer Nature format; an IEEEtran conversion script is provided under scripts/reporting/). If you use this software, please cite it via CITATION.cff (Zenodo DOI: 10.5281/zenodo.19147649).
BSD-3-Clause. Benchmark datasets remain subject to their original licenses.