Skip to content

Deepfake Research

Ref: https://github.com/flyingby/Awesome-Deepfake-Generation-and-Detection

Problem Definition

Deepfake Generation

Deepfake Generation tasks can essentially be expressed as controlled content generation problems under specific conditions, such as images, audio, text, specific attribute.

\[ I_o = {\bm{\phi_{G}}}(I_{t}, C) \]
  • Target image \(I_t\)
  • Condition information \(C = \{ \mathtt{Image}, \mathtt{Audio}, \mathtt{Text}, \dots \}\)
  • Generation network \(\phi_{G}\)
  • Output image \(I_o\)

Deepfake Detection

Deepfake Detection tasks can be viewed as an image-level or pixel-level classification problem.

\[ S_o = {\bm{\phi_{D}}}(I_{o}), \]
  • Detection network \(\phi_{D}\)
  • Fake score \(S_o\)

Tasks

  1. Face Swapping

    • Replacing the identity of the target face with that of the source face
    • Maintaining target-specific, ID-irrelevant attributes such as skin tone and expressions
  2. Face Reenactment

    • Transferring the facial movements from a driving image or video to a target image
    • Keeping the target's identity and attributes unchanged
    • Relying on facial motion capture techniques
  3. Talking Face Generation

    • Generating a talking video for a target image driven by audio, text, or multimodal inputs
    • Accurately reflecting the driving information, including lip motion, facial pose, emotions, and spoken content
  4. Facial Attribute Editing

    • Modifying semantic attributes of a target face (e.g., age, expression, or skin tone)
  5. Forgery Detection

    • Detecting and localizing tampering or forged regions in images or videos

Methods

Generative Framework

See AIGC for more details.

Datasets

Metrics

Peak Signal-to-Noise Ratio (PSNR)

Structured Similarity (SSIM)

Comments