Decision-Preserving Watermarking of Chest Radiographs: code, checkpoints, and evaluation output

Résumé fourni par la source

Code, model checkpoints, and archived evaluation output for the paper "Decision-Preserving Watermarking of Chest Radiographs" (S. Sadi, E. Sadi, T. Hedir, and M. Daoui, 2026). A learned embedder hides a 32-bit or 64-bit payload in a 224x224 chest radiograph at a fixed PSNR. The proposed objective penalises the shift of a frozen guide classifier's decision variables (a Fisher-weighted logit distance over its live CheXpert heads). The diagnostic-agnostic baseline is the same model trained with that term switched off. Preservation is measured on held-out classifiers that never entered any loss, as a per-pathology AUROC equivalence test with a patient-clustered bootstrap and a pre-registered margin of 0.005, McNemar decision-flip tests at a fixed operating point, and logit-shift statistics. Contents of stega-chest.zip:- src/: all training, evaluation, and statistics code (PyTorch), with pinned requirements.- checkpoints/: best.pt and the configuration snapshot of every arm reported in the paper (main arms, ablations, PSNR and payload sweeps, replicates, classical DWT+DCT and RONI baselines), and the ConvNeXt-Tiny held-out evaluator.- results/: the archived evaluation output that every number in the paper is generated from (calibration probes, noise floor, diagnostic endpoint per arm, bit-accuracy battery per arm).- tests/: self-checking tests for the statistics and the classifier wrapper.- reproduce.sh: the full pipeline, stage by stage. With the shipped checkpoints, the paper's tables regenerate without any training. The CheXpert-v1.0-small dataset is not included and must be obtained from Stanford under its own data-use agreement; datasets/README.md describes how to build the expected archive. Pretrained public classifiers (torchxrayvision) are downloaded on first use. All reported runs used one 11 GB GPU. One training arm takes about 1.7 hours; the diagnostic endpoint takes about 20 minutes per arm.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.