High level thoughts about general fake video - liyinnbw/deepfake-detector GitHub Wiki

Possible types of fake

  • Digital manipulation
  • Wrong Time (claim to be present but was filmed long ago)
  • Acting

Videos can differ on

  • Sensor
  • Frame rate (normal camera video and CCTV has very different rates)
  • Format (some are common, some are proprietary)
  • Compression method

Can fake video be used as evidence in court?

"Authentication requires a significant amount of known information. The affiant of the evidence must establish its lineage which identifies who created it, when it was created, and with what technology. “If you can’t answer those questions with supportive evidence,” Fredericks added, “the video would never be admitted in court.” - source

Often, when used in court, a video is only one piece of a puzzle and prosecutors, defense counsel or judges will ask someone to come and testify about the video and its authenticity. Ideally, this is a person who created the footage because what needs to be verified is time, location and other relevant facts about the video. A lot of it can be very technical and we might hear questions like: where did you stand when you made it (and this can be done in combination with maps and still images being shown in the courtroom); what type of equipment did you use, which direction were you facing, what time of day it was. Clearly, as technology developed a lot of this information can be obtained from the file itself (but this was not necessarily the case earlier). If the person who made the file is not available, there will often be people who are in the footage to confirm what the footage shows and speak about it. -source

This means if a fake video wants to be accepted as evidence

  • Someone must claim to have taken the video, or appeared in the video.
  • The video must not show sign of manipulation to both forensic tools and naked eyes.
  • The video must coincide with the claimed taking device/technology.
  • The video must coincide with the claimed location, direction, time, etc.

While it might be hard to manipulate the video without leaving digital trace, it is a lot easier if the whole scene is acted but taken in original form. However, something might not be easy to act, for example, a specific person (can use face swap, but the body must match), a hard to access place (for example the white house), etc.

Useful links: Using Video for Documentation and Evidence

Questions to ask

  • Video fake
    • Visual
    • Audio?
  • Fake that do harm
    • Define harm?
    • Harm in what context?
  • Fake for what purpose?
    • Court evidence
    • Social media mass influence
    • Personal attack
    • Impersonation
  • Types of detection methods:
    • by manipulation unit
      • Pixel level manipulation
      • Frame level manipulation (delete/add frames)?
      • Video level manipulation (change metadata like time/location)?
    • by availability of additional information
      • Active (has additional info, such as watermark)
      • Passive (no additional info other than the video content)
    • by ability to localize fake (not just to classify video as fake, but also tell where is the fake)
    • by generality
      • domain specific (works only for certain video content)
        • human, animal or plant?
        • moving or static object?
        • scene or object?
      • general (works for any faked video)
    • by whether deep learning was used in the fake
  • Is the shooting camera allowed to move?
  • Is the video allowed to be compressed? If so what level of compression?
  • What is the time limit constraint to analyze a video?

Literature Review

⚠️ **GitHub.com Fallback** ⚠️