• floofloof@lemmy.caOP
      link
      fedilink
      English
      arrow-up
      10
      arrow-down
      1
      ·
      3 days ago

      I heard that some of the checks require video of the person performing particular movements. So a little harder to fake with AI.

      • Cyrus Draegur@lemmy.zip
        link
        fedilink
        English
        arrow-up
        20
        ·
        3 days ago

        There are video deepfake systems that can animate these fictional faces. We need a catalogue of the kinds of facial positions these systems ask for

        • ramble81@lemmy.zip
          link
          fedilink
          English
          arrow-up
          5
          arrow-down
          3
          ·
          3 days ago

          One of them that is hard is waving your hand in front of your face. It trips up almost every deepfake system I’ve seen. That’s a common test we’ve had to start using with interviewees because they’d try to game the system.

          • golden_zealot@lemmy.ml
            link
            fedilink
            English
            arrow-up
            3
            ·
            edit-2
            3 days ago

            I guess in that case you could do the following, though it is a bit of work:

            1. Learn what actions are supposed to be taken in the demand for selfie identification

            2. Record a video of yourself performing the actions, but not with your hand moving in front of your face for that section

            3. Record a different video of just your hand moving past the camera

            4. Deepfake the face video

            5. Composite the video of the hand over top of the deep fake video with chroma/luma keying

            6. Use virtual webcam software to play back the video into the actual demand for selfie identification

            Wouldn’t work if a real human being is there asking, but should work for a known set of instructions ahead of time that is being audited by software.