Skip to content
GUIDEUnderstanding fake news, misinformation & disinformation

DeepForensics-1.0

Get the data

Large-scale dataset for face forgery detection research containing 60,000 videos with 17.6 million frames, including 50,000 real source videos and 10,000 manipulated videos with systematic perturbations.

Dataset characteristics

  • Total videos: 60,000 videos (50,000 real, 10,000 manipulated)
  • Total frames: 17.6 million frames
  • Real source videos: 1,000 curated YouTube videos (12.6M frames)
  • Manipulated videos: 10,000 generated via DF-VAE (5.0M frames)
  • Actors: 100 paid professional actors (95 males, 45 females aged 20-45 years from 28 countries)
  • Video resolution: 1920 × 1080 fps (HD quality)
  • Controlled conditions: Professional indoor studio environment with seven HD cameras at different angles

Key features

Controlled source collection:

  • High-quality face capture under professional lighting with seven camera angles
  • Diverse identity representation: 55% male, 45% female; 28 different countries represented
  • Eight facial expressions per actor: neutral, angry, happy, sad, surprise, contempt, disgust, fear
  • Nine lighting conditions from different directions
  • Clean faces without occlusion or glasses

DeepFake Variational Auto-Encoder (DF-VAE):

  • Novel end-to-end face swapping framework consisting of:
  • Structure extraction module for face reenactment
  • Disentangled module separating appearance from structure
  • Fusion module with temporal consistency constraints
  • Addresses artifacts in prior datasets through structure-appearance separation
  • MAdaIN (Masked Adaptive Instance Normalization) for style matching
  • Temporal continuity constraints to avoid frame discontinuities

Systematic perturbations: Seven types of distortions applied at five intensity levels:

  1. Color saturation change
  2. Local block-wise distortion
  3. Color contrast change
  4. Gaussian blur
  5. White Gaussian noise in color components
  6. JPEG compression
  7. Video compression rate change
  • Total: 5,000 videos with single-level distortions + 1,000 with random-level distortions + 3,000 with random-type distortions

Comprehensive evaluation:

  • Hidden test set with manipulated videos from unknown methods
  • User study (N=100 actors) validates realism: average rating 4.7/5.0 (vs. 4.5 for FaceForensics++, 2.5 for Celeb-DF)
  • Evaluation of five representative detection methods: C3D, TSN, I3D, ResNet-LSTM, XceptionNet

Data split

  • Standard training/test split: 1,000 real manipulated videos in training set with ratio 7:1:2 (training:validation:test)
  • Test set: 1,000 YouTube videos from the standard set (diverse identities and scenarios)
  • Hidden test set: 1,000 additionally selected videos with unknown manipulation methods

Papers using this dataset

Deepfakes & multimodalDetection methods DeepForensics-1.0: A Large-Scale Dataset for Real-World Face Forgery Detection Jiang et al. · arXiv · 2020