paper web signal

TwinD deepfake detector hits 96.3% AUROC on 19-dataset test

TL;DR

  • TwinD reports 96.3% mean AUROC across 19 deepfake datasets from 2019-2026, 8.9 points above the prior state-of-the-art.
  • Prior methods scored 96.2% on blending fakes but dropped to 78.1% AUROC on entire face synthesis datasets.
  • Models trained only on synthetic 'twins' of real video, with no real deepfakes, still average 94.3% AUROC.

TwinD, a deepfake detector from Andrii Yermakov, Jan Cech, Mario Fritz and Jiri Matas, reports 96.3% mean AUROC across 19 datasets released between 2019 and 2026, 8.9 points above the prior state-of-the-art, according to the arXiv preprint.

The gap the paper sets out to close sits in the numbers. Prior methods managed 96.2% AUROC on so-called blending fakes, where a synthetic face is swapped in over an original frame, but collapsed to 78.1% on entire face synthesis, where a generator produces the full face from scratch. TwinD holds 98.1% on blending and 95.6% on entire face synthesis.

To get there, the authors train on two kinds of synthetic 'twins' of real video rather than on any real deepfakes. A reconstruction twin runs real frames through an autoencoder and then swaps random regions back to the original pixels; a blob twin blends altered frame regions into the original, 'requiring no generator or face masks.' Models trained only on these twins still average 94.3% AUROC.

'Code, trained model, and training data are available,' the paper states, under a CC BY 4.0 license. The abstract publishes no per-detector numbers across the 19 datasets, and the stated breakdown of ten blending plus five entire face synthesis benchmarks does not fully account for the remaining four.