FakeI2V-Bench

Datasets

The following datasets are used in our benchmark:

• CelebDF-V2
• DeepFakeDetection
• GenVideo
• GenVidBench

They are all open-source and can be publicly downloaded.

Detection Methods

We evaluated the following detectors:

• Image-level detectors: 12 methods (original code available CNNDet, LNP, Patch, LGrad, DIRE, DeFake, CoDE, RINE, UniFD, DRCT, DMID, NPR)
• Video-level detectors: 8 methods (original code available FTCN, AltFreezing, UIA-ViT, LAA-Net, DeMamba, VGMShield, MM-Det, D3 )

Directory Structure

FakeI2VBench
├── Bench
│ ├── Image-level
│ │ ├── CNNDet
│ │ ├── DeFake
│ │ └── ...
│ └── Video-level
│ ├── D3
│ ├── VGMshield
│ └── ...
│
├── IV-Bridge
│ ├── VFT
│ └── MMA

Description

• Bench/
  Contains code of deepfake detectors on four datasets.

  • Image-level/ — Code for 12 image-level detectors, e.g., CNNDet, DeFake, etc.
  • Video-level/ — Code for 8 video-level detectors, e.g., D3, VGMshield, etc.

• IV-Bridge/
  Contains methods for improving detection performance.

  • VFT/ — Used to train 12 image-level detectors.
  • MMA/ — Implements multi-mode aggregation using Random Forest.