Are blood-flow biometrics the answer to video deepfakes?
A team of researchers say they have used a biometric technique to uncover deepfakes and determine the generators that created them.
Spotting deepfakes in this case means searching digitally captured faces for evidence of a pulse, a biometric signal which is hard to fake. The team says that FakeCatcher, as they call their AI system, managed a deepfake-detection success rate of 97.29 percent in portrait videos.
The two researchers from Binghamton University and a third from Intel Corp. claim in an open-source paper to use an existing technique that amplifies the effects of skin-deep blood flow.
A 2012 video unrelated to this new research demonstrates what blood flow amplification looks like on a live person’s face, and it is disconcerting, to say the least. A subject’s face rapidly flashes waves of waxy yellow to deep burgundy and back, not unlike a cuttlefish’s biological signals.
The researchers say that creators of deepfakes have not yet been able to mimic this effect convincingly.
Their source-detection approach “achieves the prediction for authenticity of the video by 97.29%, and the generative model by 93.39% on the FaceForensics++ data set.” They found that the projection of generative noise into biological signal space can create unique signatures per model,” which aids in identifying a deepfake’s generator.
Blood flow has been researched as a method for detecting biometric spoof attacks, and was even reported to be implemented in the Samsung Galaxy S10’s fingerprint sensor, though the feature’s effectiveness was reportedly quite limited.