How do you know the face on a video call is real? Measured numbers from a replay attack

Fact 1: An attacker does not need your password to impersonate you; they only need a recording of your face.

Fact 2: In testing, a replayed video of an unfamiliar person was identified with a similarity score of 0.09, far below the acceptance threshold of 0.45.

Fact 3: A printed photo of the correct person still passed the identity check, while a printed photo of the enrolled person matched at a 0.53-0.56 similarity score, above the acceptance threshold.

Fact 4: A screen playing a recorded video of a live session passed a liveness gate 39.5% of the time, despite the recording containing challenge answers on loop.

Fact 5: The timing of the challenge greatly limits this attack; the system only gives the caller 0.3 seconds to respond to a randomly chosen challenge.

Fact 6: No amount of image quality measurements, such as sharpness, noise, or moire, could differentiate a replayed video from a live face.

Fact 7: The only way to effectively detect a replay attack is by using a physical signal, such as a random color cast or depth/IR hardware, which a replayed video cannot replicate.

Fact 8: To verify people accurately, separate identity matching from liveness checks and implement random, timed challenges with a sub-second reaction window.

Fact 9: If a vendor claims their system can detect screen attacks through image quality alone, ask for their replay-attack pass rate, not their accuracy on still images.

Fact 10: The percentage of a looped screen replay of a cooperative live session that passes the gate should be publicly disclosed by every vendor.