TwinD hits 95.6% AUROC on face synthesis deepfakes, beating prior detectors
Andrii Yermakov · hf · 2026-11-20
TwinD is a deepfake detector trained on synthetic "twins" derived from real frames, generalizing across both blending and entire-face-synthesis forgeries.
- Training: builds reconstruction twins (autoencoder reconstruction with random grid/noise masks) and blob twins (color/contrast alterations with blob masks) from real frames; a frozen PEcore L backbone is fine-tuned with LoRA, with paired images sharing augmentations.
- Results: mean AUROC of 95.6% across 5 EFS datasets (vs 73.8% for AIGC detectors and 78.1% for facial deepfake detectors) and 98.1% across 10 blending datasets (baselines: 71.4% / 96.2%).
More from Research
- Team claims Codex cracked 35 open Erdős problems in 2 months, some Lean-verified — IgorCarron · 2026-11-20
- ForgeSystem paper proposes AI agents forging demand-specific software — Div_pradeep · 2026-11-20
- Accord: open-source dense metric mapping from a single camera and IMU — BE2R · 2026-11-20
- AltRA-Test: ~100 samples suffice to decide if LLM judges can replace humans — Technion · 2026-11-20
- Watch a single artificial neuron compute: weights, bias, and one activation function — TinfoilTricorn · 2026-11-20
- Frostbite: A new CPU architecture discovered, designed and verified by AI — shioyama · 2026-11-20