Prevalence-Conditioned Risk-Limiting Authorization of Deepfake Video Detectors under Dataset Shift: Aggregate Results and Reproducibility Code
Year: 2026
Type: dataset
Abstract: Aggregate results and reproducibility code for the manuscript “Risk-Limiting Operating-Point Authorization for Deepfake Video Detectors under Benchmark Shift: Identity-Isolated Certification and Modern-Model Stress Testing.” Version 6.5 adds the outcome-blind FakeSTormer extension and its independent audit to the earlier FakeAVCeleb, DF40, VCF, simulation, and multi-architecture evidence. Label-blind inference covered 21,544 FakeAVCeleb videos and 800 VCF videos. A pre-outcome statistical amendment separated 350 development, 75 independent certification, and 75 evaluation identities; no tier had a feasible development candidate, so certification and confirmation were not opened and no threshold was repaired. The release includes project-authored code, aggregate outputs, privacy-sanitized configuration derivatives, exact original-record and prediction hashes, amendment and correction records, and independent audits. It excludes source media, credentials, private paths, identities, crops, row-level labels, per-item predictions, and third-party weights. Documentation and aggregate artifacts are CC BY-NC 4.0; project-authored Python and PowerShell code is MIT.
Authors: Li, Baocai, Wang, Chao, Wang, Guangbi, Shi, Jinyu
Institution: Baoding University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7202030002