Deepfake Detection Based on Original Human Biometric Traits
... Successful and popular deepfake methods such as FaceSwap and DeepFaceLab/Livecurrently have zero capacity to create such granular biometric approximations, relying at best on talented impersonators on whom the faked identity is imposed, and much more commonly on apposite in-the-wild footage of ‘similar’ people. ...
These two dominant deepfake packages are based on autoencoders. Alternative human synthesis methods can use a Generative Adversarial Network (GAN) or Neural Radiance Field (NeRF) approach to recreating human identity; but both these lines of research have years of work ahead even to produce fully photorealistic human video.
With the exception of audio (faked voices), biometric simulation is very far down the list of challenges facing human image synthesis. In any case, reproducing the timbre and other qualities of the human voice does not reproduce its eccentricities and ‘tells’, or the way that the real subject uses semantic construction. Therefore even the perfection of AI-generated voice simulation does not solve the potential firewall of biometric authenticity. ...
See the full lengthy story here: https://www.unite.ai/deepfake-detection-based-on-original-human-biometric-traits/
