Large-scale deepfake detection dataset containing 5,639 high-quality DeepFake videos generated from 590 real videos of 59 celebrities using an improved face-swapping synthesis algorithm.
Dataset characteristics
- Total videos: 5,639 DeepFake videos + 590 real videos
- Total frames: >2 million video frames
- Video length: ~13 seconds at 30 fps
- Format: MPEG4.0
- Subjects: 59 celebrities (56.8% male, 43.2% female) with diverse ages, ethnicities, and backgrounds
- Visual quality: High-quality synthesized faces (256×256 resolution) with significantly fewer artifacts than prior deepfake datasets
Key features
Improved synthesis algorithm: Addresses five key artifact categories in prior datasets:
- High-resolution face synthesis (256×256 vs. 64×64 or 128×128)
- Color correction via data augmentation and color transfer
- Better face masks with surrounding context
- Temporal smoothing via Kalman filtering
- Boundary smoothing between synthetic and original regions
Comprehensive evaluation: Includes evaluation of 13 deepfake detection methods (Two-stream, MesoNet, HeadPose, FWA, VA, Xception, Multi-task, Capsule, DSP-FWA) across 6 datasets.
Related papers
- 2019-li-celeb-df-dataset — Introduces the Celeb-DF dataset and comprehensive detection evaluation
- 2020-dolhansky-dfdc-dataset — Concurrent second-generation dataset (DFDC) with consenting actors
- 2019-rossler-faceforensics-plus-plus — FaceForensics++ dataset (first-generation benchmark)
Related topics
Papers using this dataset
Deepfakes & multimodalDetection methods
Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection
Deepfakes & multimodalDatasets & evaluation
Deepfake Detection: A Systematic Literature Review
Deepfakes & multimodalDetection methods
Celeb-DF: A Large-scale Challenging Dataset for DeepFake Forensics
Deepfakes & multimodalDetection methods
Face X-ray for More General Face Forgery Detection