[ 01 ]
What gives a face swap away
A swapped face is generated separately and blended into the original frame. The seam can show brightness flicker, edge softness or lighting that does not follow the scene, particularly when the head turns or something crosses the face.
[ 02 ]
Measurements used
Temporal-consistency sensors track the face region frame by frame. Noise-residual sensors compare the face with the background, because a pasted region rarely shares the camera sensor's noise. Compression sensors look for a face region that was encoded differently from its surroundings.
[ 03 ]
Lip-sync and re-animation
Lip-sync deepfakes change only the mouth to match new audio. Mouth-region consistency checks and comparison between audio and video timing help flag them, with the limits stated in the report.
[ 04 ]
Limits
Low resolution, heavy compression and social-media re-uploads weaken every sensor. A responsible result records confidence and does not claim certainty.
Quick answers
Can face swaps be detected in compressed video?
Often partly. Compression reduces evidence, so SpearTrace states how reliable each sensor is for the file in front of it.
Is detection done live?
SpearTrace analyses recorded files and returns results after analysis rather than screening live streams.