Fast deepfake detectors were built for older generators and lose accuracy on the newest ones. I'm working on closing that gap without giving up speed on a laptop.
How it works, roughly
- A fast base detector handles the common case. It is what keeps inference cheap enough for a laptop.
- Two extra views catch what the base misses: frequency-domain artifacts from upsampling and generation, and a CLIP embedding that looks at the image semantically rather than pixel by pixel.
- A small gating layer learns how much to trust each view for a given image. Only that layer is trained; the rest stays frozen, so experiments run in seconds.
- The target is generalization: the detector is judged on generators it has never seen, not on the ones in the training set.
Results and details will follow when I can share them.