thanatchon36/beyond99-banknote-detection

Beyond 99%: open-set evaluation of counterfeit banknote detectors

Data and results for the preprint Beyond 99%: Open-Set Evaluation Reveals Silent Failures of Counterfeit Banknote Detectors, and the Case for One-Class Screening by Thanatchon Chongmankhong (read online · PDF). The paper evaluates counterfeit detectors on counterfeit families they have never seen, at a fixed 1% genuine-rejection budget. Under that protocol, tree ensembles that score above 99% on random splits catch 51–61% of unseen-family counterfeits, while a one-class autoencoder trained on genuine notes alone catches 99.98%. Code: https://github.com/thanatchon36/beyond99-banknote-detection; trained detectors: thanatchon36/beyond99-banknote-detection; archived on Zenodo: 10.5281/zenodo.24049790.

Independent personal research. Does not represent the views of any employer or institution.

Files

path	content
data/banknote.csv	UCI Banknote Authentication with a header: note_id (0–1371, row order of the UCI file), the four features variance, skewness, kurtosis, entropy of a wavelet-transformed image, and label (0 = genuine, 762 notes; 1 = counterfeit, 610 notes)
data/data_banknote_authentication.txt	the original UCI file, unchanged (no header; SHA-256 in SHA256SUMS)
data/counterfeit_families.csv	the counterfeit "families" used by the paper: family_k2 … family_k6 give each counterfeit's K-means cluster (standardised features, n_init=20, random_state=0) and −1 for genuine notes. Family sizes at K = 3: 71, 360, 179
results/e1_folds.csv, e1_summary.csv, e1_test.csv	E1, random-split benchmark: 10 supervised models, repeated stratified 10-fold cross-validation (5 × 10); per fold, summary, and the corrected resampled t-test of the best tree ensemble against the MLP
results/e2_learning_curve.csv	E2, accuracy and miss rate against 10–640 training notes (30 repetitions)
results/e3_ablation.csv	E3, a weaker sensor: all four features, without variance, kurtosis and entropy only
results/e4_noise.csv	E4, Gaussian noise on test features (worn notes, sensor drift)
results/e5_unseen_families.csv, e5_summary.csv	E5, unseen counterfeit families at K = 3: per seed and setting (family 0..2 withheld, or random split), and the summary
results/e5x_K2.csv … e5x_K6.csv	E6, leave-one-family-out for K = 2..6: 10 seeds × (K families + random split) × 18 detectors × 7 budgets alpha ∈ {0.2%, 0.5%, 1%, 2%, 5%, 10%, 20%}. cdr = counterfeits detected in the held-out family, grr = genuine test notes rejected, auc = ROC AUC (empty for hybrids). e5x_K3.csv at alpha = 0.01 is Table 5 of the paper
results/quadrant.csv	mechanism analysis, K = 3, 1% budget: for every genuine test note and every held-out counterfeit (idx = note_id), the exact margin of XGBoost or the MLP (sup, threshold sup_thr, logit units), the squared Mahalanobis distance from the genuine class (maha, maha_thr) and the autoencoder reconstruction error (ae, ae_thr)
SHA256SUMS	checksums of every data and result file

Detectors: supervised (logistic regression, k-NN, SVM with RBF kernel, decision tree, random forest, extra trees, HistGradientBoosting, XGBoost, LightGBM, MLP 3 × 64), one-class (autoencoder, Mahalanobis, one-class SVM, isolation forest) and hybrids that flag a note if either component exceeds its own α/2 threshold. Thresholds are the (1 − α) quantile of each detector's scores on held-out genuine notes. Every number regenerates with make all in the code repository.

Usage

from datasets import load_dataset

notes = load_dataset("thanatchon36/beyond99-banknote-detection", "banknote", split="train").to_pandas()
lofo = load_dataset("thanatchon36/beyond99-banknote-detection", "lofo_k3", split="train").to_pandas()
t5 = lofo[(lofo.alpha == 0.01) & lofo.setting.str.startswith("family")]
print(t5.groupby("model").cdr.mean().sort_values())   # mean unseen-family detection, Table 5

Source and licence

The banknote data are the UCI Banknote Authentication dataset (Lohweg, 2013, doi:10.24432/C55P57), released by the UCI Machine Learning Repository under CC BY 4.0 and redistributed here unchanged apart from the added header and note_id. Images of genuine and forged banknote-like specimens were taken with an industrial print-inspection camera (400 × 400 pixels, about 660 dpi) and reduced to four wavelet statistics. The family assignments and all results are by Thanatchon Chongmankhong, also CC BY 4.0. Counterfeit "families" are K-means clusters, a statistical proxy for real production processes, not labels from the data source.

Citation

@misc{chongmankhong2026beyond,
  author = {Chongmankhong, Thanatchon},
  title  = {Beyond 99\%: Open-Set Evaluation Reveals Silent Failures of Counterfeit Banknote Detectors,
            and the Case for One-Class Screening},
  year   = {2026},
  note   = {Preprint},
  doi    = {10.5281/zenodo.24049790},
  url    = {https://thanatchon36.github.io/beyond99-banknote-detection/paper/}
}

@misc{lohweg2013banknote,
  author       = {Lohweg, Volker},
  title        = {Banknote Authentication},
  howpublished = {UCI Machine Learning Repository},
  year         = {2013},
  doi          = {10.24432/C55P57}
}