Local-Fake Knee Radiograph Dataset and Detector Artifacts for Cross-Source Deepfake Detection (v1)

his dataset package supports a study on synthetic-image forensics in knee radiographs.

Research hypothesis
A detector trained only on a local real/fake development domain can learn source-robust fake-vs-real ranking features, but a decision threshold locked on local validation may not transfer perfectly to external real-image domains without site calibration.

What this package contains

A public Local-Fake knee radiograph set generated from a local retrospective cohort (local real data are not publicly released due to privacy constraints).
Image-level manifests linking image IDs, metadata, and file paths.
Two retained ResNet50 detector artifacts and their evaluation outputs:
Primary zero-shot model: local_roi_cutoutcrop_pseudofake_p10_smallpatch_softblend_seed13
Secondary comparator model: local_roi_cutoutcrop_deg1_pseudofake_seed13
Reproducibility files (checksums, configs, threshold-lock metadata, model cards).
How data were generated and curated

Local fake images were synthesized from the local development subset using a conditional StyleGAN3 pipeline, with one accepted synthetic image per source image after QC.
Initial generated pool size: 3366 images.
After manual adjudication of near-duplicate candidates, 6 images were excluded.
Final released Local-Fake count: 3360.
QC summary for released pool: exact duplicates to source = 0; exact duplicates within fake set = 0.
The explicit exclusion list is provided in manifests/local_fake_excluded_near_duplicate_6.csv.
Notable findings from included model evaluations

Primary zero-shot model (locked local threshold): strong fake sensitivity on external synthetic sets (Pitt and Prezja) with moderate external real specificity on OAI.
Secondary comparator: stronger cross-source ranking AUROC but lower external real specificity at the locked local threshold, consistent with threshold-transfer/calibration shift.
How to interpret and use this dataset

Use manifests/local_fake_public_manifest.csv and manifests/local_fake_public_file_index.csv to map files to IDs and metadata.
Use the included model cards and threshold-lock files to reproduce reported operating points.
Interpret this package as a research benchmark for synthetic-image detection robustness and calibration behavior, not as a clinical diagnostic tool.
External datasets (OAI real, Pitt fake, Prezja fake) are referenced in the study but must be obtained from their original public sources and licenses.
License and governance

Package license: CC BY 4.0 (see LICENSE_DATA.md).
Local real radiographs are not released due to privacy and institutional governance constraints.