quartz/content/notes/19-deepfakes.md

title	aliases	tags
19-deepfakes		comp210 lecture

what, origins, how

• issue with trust politics etc

• essentially better face swapping

• replaces traditional studio artist with ML

• open source

  • easy to do
  • point and click
  • more widely available

• in early stages so quality varies

related concepts

• voice faking c.f., impersonation

• lip syncing (replaces mouth no whole face)

  • take actual footage of e.g., politician and change words
  • can be done using only text
  • natural language processing

• uncanny valley

  • 
  • "something not quite right"
  • some people dont realise

how

• auto encoders

  • take image an reduce to a base representation
  • e.g., of the face

• decorder

  • takes base and produces image

• deepfake

  • takes base of two differnet people and combines them
  • two decoders
    • one for each person
      • a pathway and b pathway
    • both decode from the same combined base representation

• to swap

  • 

• deep face lab

  • ml workflow

other requiremtns

• need a lot of data
  • video - extract individual frames
• similar facial structures
• source content cleanup
  • remove images with two or more faces
  • crop images
  • resample/distort/skew image to create new instances
• post processing tools to touch up afterwards

concerns and solutions

• genuine concern about video being distict

  • we expect images and text to be fake
  • do not expect videos to be fake
    • more damaging

• politically weaponised

  • e.g., deep fakes of biden declare war on russia

plausible deniability

• porn
  • swap actors onto porn scenes
  • hard to deny
  • how to prove its a deepfake
  • revenge porn
• lack of trust
  • only from the existense of deepfakes
  • true things become more suspicious

should "you" be concerned

• for your personal integrity no
• likely not enough data for you to be faked