Arab-DF: A Large-Scale Arabic Deepfake Dataset

Arab-DF is a large-scale deepfake dataset dedicated to monolingual Arabic visual content,built from publicly broadcast television and news material. CONTENT- 1,066 authentic videos- 11,240 manipulated videos- 12,306 videos in total- 450 source identities across 22 Arab countries MANIPULATION FAMILIES1. Face swapping (FaceSwap, DeepFaceLab, FSGAN)2. Facial reenactment (First Order Motion Model)3. Lip synchronization (Wav2Lip)4. Mobile-application manipulations (Reface, DiffSnap, WOMBO, MyHeritage, FaceApp) SPLITSTraining, validation, and test splits are source-video-disjoint: all manipulated versionsof a given authentic source video are kept within a single split. The guarantee isestablished at the level of source videos, not of identities. THIS DEPOSIT CONTAINS- Dataset metadata (video identifiers, manipulation family, generator, country)- Evaluation splits- Evaluation code and benchmark protocol- Per-video detector predictions for the eleven benchmarked detectors ACCESSRaw videos are provided under restricted access due to privacy and copyright constraintsof the source broadcast material. Researchers may request access by agreeing to a datausage agreement restricting use to non-commercial forgery-detection research andprohibiting redistribution.